Changelog

August 12, 2026

AI Knowledge Builder

• Turn past tickets into a searchable knowledge base. Point the new AI Knowledge Builder source at your resolved Zendesk tickets and every ticket either becomes a new KB article or gets merged into a similar one already there. The result is a normal Macha knowledge source, chunked and embedded, reachable via search_knowledge the same way uploaded PDFs are.
• Two writing modes. Prompt mode does one LLM call per record: fast, cheap, deterministic. Agent mode picks one of your Macha agents to investigate each record first (with write tools stripped for safety), then writes a richer article using the context the agent gathered.
• Similarity-based deduplication. Before creating an article, the builder searches for semantically close existing articles on the same source. If it finds one, a small LLM decides whether to create fresh, rewrite the existing one, or attach the ticket to it without changing the body. No more duplicate articles for the same question worded slightly differently.
• Skip filter on by default. A relevance pass drops auto-replies, spam, and thin threads before the article-gen model spends credits on them. The rules are editable so your org can add its own exclusions.
• Editable articles with a rich text editor. Every generated article is marked editable. Open one from the Sources drawer or the builder's Articles tab and you land in a full-screen editor built on the same TipTap surface used for agent instructions. Save re-embeds automatically so search stays current.
• Also asked as. Each article carries a list of alternate phrasings of the question it answers. It's set by the article-gen LLM using structured output and rendered inline in the article body, so retrieval matches on the original question and its rewordings both.
• Article generation runs on GPT-5.4 and is billed per record. The skip filter and merge-decision passes run on GPT-5.4 Mini and are Macha-funded. You see the credit estimate before every batch. Available on Professional and Enterprise. Full doc at AI Knowledge Builder.

Agent Simulations

• Dry-run agents on real records before shipping. The new Simulations tab on every agent lets you replay past Zendesk tickets through the agent with write tools stubbed. New conversations get generated, tools get called, but replies never land on real tickets. The closest thing to a production preview.
• Distinct from Evaluation and Studies. Evaluation grades what already happened; Studies extract structured fields from records; Simulations run the agent through fresh records so you can see what would happen. Reach for it after a prompt change, when rolling an agent out to a new customer, or when sanity-checking a tool wiring.
• Per-record verdict, full transcript, stubbed tool trace. Every simulated conversation lands with a StudyResult verdict (Instructions followed, resolution, sentiment, category) using the same grader Auto-Eval uses. Click any row to see the reasoning trace, the tool calls the agent made, and the canned responses the stubbed writes returned.
• Snapshots frozen at run time. The adapter config and the agent's model are frozen when the run launches. Later prompt or model edits don't invalidate old results. Simulated conversations are filtered out of the main agent-evaluation view, so your production adherence numbers stay clean.
• Available on Professional and Enterprise. Each simulated conversation costs one agent run at the model's per-conversation rate plus the auto-eval judge credit. Full doc at Agent Simulations.

July 27, 2026

Two new OpenAI models: GPT-5.6 Terra and Luna

• GPT-5.6 Terra is Macha's new flagship. Same 3-credits-per-ticket rate as GPT-5.4, roughly 1M-token context window, sharper on multi-step reasoning and tool selection. If you want the best-quality answer on every ticket, this is now the default recommendation.
• GPT-5.6 Luna is the new standard mid-tier. 2 credits per ticket, same 1M-token context, faster than Terra with only a small step down in quality. Best pick for high-volume conversations where cost-per-ticket matters more than the last few percent of answer quality.
• Both are available on every plan. Open any agent, hit the model dropdown, and switch. Existing agents keep whichever model they were on until you change it. Nothing gets migrated behind your back.
• GPT-5.4, GPT-5, and GPT-5.4 Mini stay in the picker as previous-generation options. Pricing math on the pricing page and in the credits tips now uses Terra as the reference model.

July 27, 2026

Two new OpenAI models: GPT-5.6 Terra and Luna

• GPT-5.6 Terra is Macha's new flagship. Same 3-credits-per-ticket rate as GPT-5.4, roughly 1M-token context window, sharper on multi-step reasoning and tool selection. If you want the best-quality answer on every ticket, this is now the default recommendation.
• GPT-5.6 Luna is the new standard mid-tier. 2 credits per ticket, same 1M-token context, faster than Terra with only a small step down in quality. Best pick for high-volume conversations where cost-per-ticket matters more than the last few percent of answer quality.
• Both are available on every plan. Open any agent, hit the model dropdown, and switch. Existing agents keep whichever model they were on until you change it. Nothing gets migrated behind your back.
• GPT-5.4, GPT-5, and GPT-5.4 Mini stay in the picker as previous-generation options. Pricing math on the pricing page and in the credits tips now uses Terra as the reference model.

July 22, 2026

Memory — persistent notes for your general chat

• Memory is now built into the general Macha chat. Tell Macha "remember I prefer German-formal replies" or "note that our escalation channel is #support-legal," and it saves that as a memory. Every future general chat picks it up automatically without you re-explaining. Scoped agents (@handle mentions, the Chrome extension, the Zendesk widget) don't share your memory — those already have their own instructions doing the same job.
• A manage page in Settings. Settings → Memory is your control panel. Toggle memory on or off (nothing is deleted while it's off — flip it back and everything returns), open Manage to browse every saved memory, add new ones by hand, or edit and delete existing entries. Each memory holds a title, a one-line description (what the chat sees in its always-on index), and a full body (only pulled in when the chat decides it's relevant).
• A pill in every chat header telling you the state. Memory on when the general chat is using it; Memory 12/15 as you approach the cap; Memory full 15/15 when new saves are refused; Memory paused when you've toggled it off; No memory here on agent-scoped or widget chats. Every state has a tooltip explaining exactly what it means.
• Explicit-only saves. Macha only writes to memory when you ask directly. It won't decide on its own to remember things — proposal-based memory ("want me to remember that?") is on the roadmap but not on by default. Every save fires a subtle black toast confirming the write.
• Strictly per user, per organization. If you're in three Macha workspaces you have three separate memory stores. Teammates never see your memory, admins never see your memory — it's yours alone in each org. Memory is not billable and doesn't consume credits.
• Capacity limits, with warnings before you hit them. Fifteen memories per user, per organization. Title 200 chars, description 400, full body 10 000. The manage page shows a live character counter next to every field, amber at 80% of the limit and red at the cap. See the memory docs for the full walk-through.

July 21, 2026

Agent Evaluation, now continuous and free on every plan

• Unlimited auto-evaluation, included on every plan. Every conversation your agent handles in Agent mode is now graded automatically against its own instructions — no cap, no credit cost, on trial, starter, professional, and enterprise alike. Previously auto-eval stopped at 50 conversations per agent; now it just keeps running for as long as the agent runs.
• Live adherence dashboard on every agent. Open any agent's Agent Evaluation tab and you'll see the current adherence score, the Yes / Partially / No breakdown, and a trend chart showing how the score has been moving day to day. A dashed volume line overlaid on the trend tells you at a glance how many conversations were judged each day.
• AI-written notes every 50 conversations. A rolling Notes card summarises the last 50 graded conversations — the dominant pattern the agent got right, where it slipped, and 2–4 concrete instruction tweaks that would lift the score. Older notes stay browsable via a View all notes modal so you can compare batches over time and see whether a prompt change actually moved the needle.
• Resolution and sentiment on every graded conversation. Alongside Instructions followed, the AI judge now also scores whether the conversation was actually resolved (Resolved / Partially / Unresolved / Escalated to human) and how the customer sounded (Positive / Neutral / Negative / Frustrated). Both show up as their own breakdown cards, so you can see the shape of the outcome, not just the shape of the compliance.
• Every graded conversation now carries a category. Refunds, Login issues, Shipping delay, Order status, and so on — a short topic label per conversation so you can see at a glance what your agents are actually spending their time on.
• Multi-turn tickets get multi-turn grades. When a customer replies to a ticket and the agent handles that new turn, the new turn gets its own grade. You can now see how the agent held up across every fire on a ticket, not just the first one.
• Only Agent-mode conversations are graded. Auto-evaluation now focuses exclusively on the trigger-fired (autonomous) conversations customers actually deploy in production. Chat sessions in the dashboard, widget conversations, and test runs are excluded, and a labelled pill in the header spells this out. Switch the agent to Chat mode or deactivate it and the header pill flips to Auto-evaluation paused with a tooltip explaining why.
• History + Analytics, now one tab. The old Analytics tab has been renamed to History and now shows the conversation list and the tool / model / avg-time / connector context that used to live on Analytics — with a hover View tools affordance to open the tool-activity drawer without leaving the row. Row click opens the conversation inside the agent's Try it tab, so you stay in context.

July 14, 2026

Macha lives inside every Zendesk ticket now

• The Zendesk sidebar widget is live. Install Macha from your Zendesk account and the copilot shows up in the right-hand sidebar of every ticket, no separate tab, no context switching, no pasting the ticket into a chat window. Open a ticket and Macha is already on it: subject, requester, current status, assignee, tags, custom fields, and the full public / internal thread are loaded into context the moment the sidebar renders.
• Per-ticket memory that survives revisits. Every ticket keeps its own conversation. Come back to the same ticket a week later, the thread is right where you left it, not a new chat, not a search. Each ticket is a separate thread, scoped to the agent who's been working it.
• History dropdown in the header. Click the ticket subject pill at the top of the widget to see every past Macha conversation for the current ticket, newest first. Great for a ticket that's been reassigned or that you're picking up mid-flight, you can see exactly what the previous agent asked Macha and what came back.
• "+ New chat" starts a clean thread on the same ticket without losing the old one. Every conversation is preserved in the history dropdown.
• Welcome message on install. The first agent to open the widget after connecting Zendesk gets a one-time greeting oriented around the ticket they're on: a short summary and 2–3 concrete things they can ask Macha to do (draft a reply, check the customer's history, summarize the thread). One welcome per account, not per ticket, everyone else on the team just lands directly on the chat.
• Same agents, same tools, same rules. Whatever agents you've configured for Zendesk in the Macha dashboard are the same agents available in the widget. No new configuration, no separate model selection, no separate trial credits.

Trial is now credit-bound, not time-bound, 500 credits, no expiry

• 500 credits, however long you want to spend them. The old 7-day trial timer is gone. Every new Macha account starts with 500 trial credits, and the trial stays active for as long as those credits last. If you use 30 credits your first week, you have 470 credits and no deadline. If you take a month off and come back, the credits are still there.
• Upgrade prompt fires on runway, not on a calendar. No more "trial ends in 3 days" nudges when you're actively evaluating. Instead, when your credit meter runs low, you'll get a heads-up in the dashboard and an inline "add credits" button on the top banner that takes you straight to the plan picker. Runs out = upgrade. Not yet = keep going.
• Existing trials were migrated automatically. Every trial account that had fewer than 500 credits was bumped up to 500 the day this change shipped. No action needed on your part.

Credits are now charged per conversation, not per message

• One conversation = one credit deduction. Previously every user message consumed credits at the model's rate, a 100-turn conversation on GPT-5.4 could burn hundreds of credits before you noticed. Now Macha deducts credits exactly once per conversation, at the model's rate the moment the first assistant reply lands. A 3-turn discovery chat costs the same as a 100-turn deep dive. Plan your usage against ticket count, not message count.
• Model rate locks in at the first reply. If you start a conversation on GPT-5.4-mini and later switch to Claude Sonnet mid-thread, you're not re-charged, the conversation's cost is fixed at the first-reply model rate. No stacking, no surprise upgrades. If you want a different model's cost, start a new conversation.
• Sub-agent handoffs stay covered. When one of your agents calls another agent to handle a step, that sub-agent run rides on the parent conversation's single credit. No double-billing when an agent hands off internally.
• Welcome and onboarding conversations stay free. The one-time Zendesk widget welcome and the initial onboarding chat you get on signup have always been free, that hasn't changed, and they're not counted in your credit balance.
• Existing conversations aren't re-billed. Every conversation that had already consumed credits under the per-message model is grandfathered forward. Old chats you re-open won't get charged a second time.

July 3, 2026

Chat mode & Autonomous mode: at-a-glance indicators on every agent

• Mode tag next to every agent. The Agents table and the agent detail header now show a coloured tag telling you exactly how the agent runs. Chat mode (blue) means the agent only responds when you chat with it, it has no triggers and will never act on its own. Autonomous mode (orange) means the agent has one or more triggers configured and will fire on its own when those events happen (new ticket, scheduled run, incoming Slack mention, etc.). A ? icon next to the tag opens a tooltip explaining the difference.
• Why we added it. New users often looked at the Enabled toggle and worried an agent might do something without them. The mode tag settles that anxiety upfront: if it says Chat mode, nothing runs unattended, period.
• Conversations column too. Right next to the mode tag, the Agents table now shows total real conversations per agent (excludes sub-agent runs, simulations, and test-run entities) so you can see at a glance which agents are actually being used.

Auto-Eval email now embeds a written summary of the run

• Diagnostic rollup, not just a score. When an agent's Auto-Eval hits its cap, the completion email now leads with a short GPT-5–written summary explaining why the agent scored what it did, the dominant failure mode, what worked, and the specific patterns the judge saw across all graded conversations. The adherence score and Yes / Partially / No breakdown are still there; the summary sits above them as the operator's TL;DR before opening the results page.
• Coaching note in the results modal. A new View summary notes button on the Auto-Eval run page opens the full structured version: Overall performance, What worked, Failure patterns, and Recommended fixes as concrete instruction / config edits. Reads like a coaching note from a reviewer who watched every conversation.

PII anonymization for conversations at rest

• Personally identifiable data is scrubbed before it hits your logs or the model context. Emails, phone numbers, names, addresses, and identifiers are redacted from stored conversation transcripts and from anything Macha ships to a model provider. The redaction runs at the boundary with pattern-based detectors, so nothing sensitive leaks into third-party inference APIs or into your Studies exports.
• Enterprise-grade detection is next. The current pass covers the common classes fast. A follow-up upgrade to Microsoft Presidio, NER-based detection with configurable entity types and richer masking / tokenization, is queued and will drop into the same pipeline.

Document360 connector

• Knowledge base as a live source. Connect a Document360 project with an API token and Macha treats it as a live knowledge source, same pattern as Notion and Google Docs. Articles aren't chunked or re-embedded; agents call document360_get_article at read-time and always see the current version. Nine tools land: search, get, list, categories, tags, and confirmation-gated create / update / publish for write flows.
• One connector per project. Because Document360 accounts have multiple projects, each connector is scoped to a single project, pick it from a dropdown after we verify your token. Connect Document360 again to reach a second project. Tools are auto-suffixed if you have multiple instances so the LLM can disambiguate.

Interactive chat gets three new capabilities

• Reasoning trace. Agents on capable models (GPT-5, GPT-5.4, Claude Sonnet 4.5) now expose their chain-of-thought in a right-side drawer while they think. A View thinking pill on each assistant reply lets you re-open the trace after the fact. Reasoning is chat-only, autonomous trigger runs and sub-agent handoffs are unaffected.
• Inline charts. Ask your agent to visualize anything and it can render a bar, line, or donut chart directly in the reply with a View Data table fallback. Same Chart.js primitives Sidekick uses.
• Ask-user picker. Agents can now pause mid-turn and ask you to pick from a short list of options (with keyboard shortcuts 1–N, arrows, Enter). Great for disambiguating between two customers, choosing which of several tickets to escalate, or picking a template. Frames the interaction as a first-class choice rather than making you re-type the whole prompt.
• Confirmation preamble discipline. Agents now announce write operations in their own words before firing the confirmation card, so you always understand what's about to happen and why. Read-only tools still run without a gate.

June 30, 2026

Agent Auto-Eval: continuous adherence scoring, free on every plan

• Every conversation graded automatically. The moment you create an agent, Macha starts evaluating every conversation it finishes (direct chat, autonomous trigger run, sub-agent handoff, embed chat) against that agent's own live instructions. No setup, no rubric to define, no credits charged. A hardcoded GPT-5 judge fills in two locked fields per conversation: Instructions followed (Yes / Partially / No) and a short Why justification.
• Lives in your existing Studies list. The Auto-Eval run shows up with a Continuous pill, an X / N counter that ticks up as conversations are graded, and an Included label where the credit cost would normally sit. Click in and you get the same results table, the same report view, the same per-conversation drawer that manual Evaluations have given you all along, just without you having to configure it.
• Free on every plan, with a cap. 50 conversations per agent, across every plan. When you hit the cap, Macha emails you the final adherence score and a breakdown of Yes / Partially / No verdicts. The Study stays in your list as a permanent record. To keep grading beyond the cap, kick off a manual Evaluation over the same scope.
• Locked from edits. Auto-Eval Studies can't be renamed, re-scoped, or deleted. That's the trade-off for it being free and uniform across every customer: the rubric is opinionated, so the headline adherence number you see is computed the same way as every other customer's.
• Same machinery as the paid Evaluations. Auto-Eval reuses the existing Studies, StudyRun, and StudyResult system end-to-end. Adherence score formula is identical: Yes + ½ × Partially, divided by total. A 60% Yes, 40% Partially run scores 80%.

June 26, 2026

Sidekick can now test your agents end-to-end

• Test-and-iterate loop. Ask Sidekick to test an agent and it drafts realistic scenarios from the agent's own instructions, walks you through a confirmation, then creates a real entity on the connected workspace (a real Zendesk ticket, a real Freshdesk ticket, a real Front conversation, a real Gorgias ticket, or a real Slack message) and dispatches your agent autonomously against it. No synthesized payloads, the agent runs the same code path it runs in production.
• Live TestRunCard. The run streams into your chat as it happens, every tool call and every assistant reply visible, with status flipping from Running to Completed (or Failed) when the agent finishes. No spinner-and-wait, you're watching it work.
• Automatic review with concrete tweaks. When the run completes, Sidekick reads the full transcript and writes a tight review in the same turn: what worked, where the agent stumbled (with the exact tool call or message quoted), and 2–3 specific instruction or tool tweaks to fix it. The picker that follows lets you apply the fix, run another scenario, or skip.
• Cross-connector webhook short-circuit. Test entities are registered the moment they're created. Other triggers in your org that watch the same event (a second Zendesk trigger on ticket.created, a Slack app-mention trigger, etc.) get the webhook dropped at the routing layer so they don't accidentally fan out to your test ticket. You can run tests on a busy connector without polluting the other agents in your workspace.
• Sidekick suggests tests at the right moments. After creating an agent, after a non-cosmetic edit (instructions, tools, or sub-agent changes), and after an audit that surfaced a concrete fix, Sidekick proactively offers to run a test. No need to remember to verify, the prompt comes at every write boundary.
• Test entities are yours to keep or delete. Sidekick won't auto-close or auto-delete the test ticket after the run. The full record sits in your connector with a clearly-fake requester email so you can inspect it side-by-side with the review.
• Available on every plan. Five connectors covered at launch: Zendesk, Freshdesk, Front, Gorgias, and Slack.

Agent detail page: Edit with Sidekick + Test run pills

• "Edit with AI" is now "Edit with Sidekick" with the Sidekick mascot icon, sitting in a rounded pill next to Test run. The two buttons are visually siblings, matching height and shape, so configuration and verification feel like one continuous flow.
• Test run pill moved to the right-side action group on the configuration page, next to Edit with Sidekick and Save. Cleaner separation between the section title on the left and the action verbs on the right.

Playbook rename for clarity

• evaluate-agent is now audit-agent-runs, same retrospective behavior (read past conversations, judge them against the agent's instructions), but the name now makes the boundary explicit. audit-agent-runs looks at history; the new test-agent playbook creates fresh runs. Two sibling workflows, each clearly named for what it does.

June 25, 2026

Public REST API and AI-ingestible docs

• API Reference is now live at /docs/api. A full browseable reference for the Macha REST API: every endpoint, its request body, query parameters, response shape, status codes, and authentication requirements, with copy-paste cURL examples on each page.
• Covers the full read + write surface: conversations (list, retrieve, full message thread), agents (list, retrieve, create, update, delete), custom tools (list, retrieve, create, update, delete), sources (list, retrieve, create, update, delete), and triggers (list, retrieve, create, update, set active, delete). Use it to embed Macha agents into your own workflows, sync agent configuration from a source of truth, or pull conversation data into your own analytics stack.
• Authentication via API keys, scoped per organization. Generate and rotate keys from Settings → API Keys. Keys carry the same plan limits as the dashboard, no extra metering.
• llms.txt for AI agents. Crawlers like Claude, ChatGPT, and Perplexity can now ingest the full Macha knowledge base via /llms.txt (short index of every doc page and API endpoint) and /llms-full.txt (full prose content concatenated), following the llmstxt.org convention.

Sidekick v2: a full in-product copilot for admins

• Full-screen workspace. Sidekick now opens as an overlay panel (75% width, 85% height) over your dashboard with a backdrop blur. Hit ⌘K (Mac) or Ctrl K (Windows/Linux) from anywhere in the app to launch it. The session keeps your place if you close the overlay and reopen it.
• Built for admins, not just chat. Use Sidekick to build new agents, edit existing ones, recommend tools for an agent, create and debug custom API tools, set up scheduled triggers, and audit how your agents are performing on real conversations. Conversation analytics live alongside agent setup, so you can evaluate and tune in the same chat.
• Reasoning trace, exposed. Every assistant turn now shows a per-message Reasoning pill. Click any narrator line (italic status updates between tool calls) to open the Reasoning drawer, with the model's full chain-of-thought streaming in alongside the chat.
• Web search built in. Sidekick can search the web when your question goes beyond what's in the Macha docs (third-party API specs, vendor pricing, current events). Inline citations appear right under the answer.
• File uploads. Click the paperclip or drag any file onto the panel: images, PDFs, text, CSV, JSON, Markdown, and DOCX up to 20MB each, 5 per message. Sidekick reads them natively and quotes back what it found.
• Voice input. Tap the microphone to dictate with Groq Whisper transcription. Auto-sends when you stop talking.
• Inline charts. Ask Sidekick to visualize anything (bar, line, or donut), say "chart credits used per agent this week" or "plot conversations per day", and a Chart.js card renders inline with a View Data fallback table.
• Claude-Code-style interactions. Confirmation cards before destructive actions, multiple-choice pickers with number-key shortcuts (1, 2, 3, ↑/↓, Enter), ghost-text suggestions you can accept with Tab, and a collapsible history rail with auto-titled sessions.

June 23, 2026

Two new helpdesk integrations, Gorgias and Front

• Gorgias, built for Shopify-native support teams. 15 tools covering tickets, messages, customers, macros, tags, and team assignment. Custom-webhook triggers for ticket created, updated, message added.
• Front, shared-inbox-first. 15 tools covering conversations, contacts, custom fields, inboxes, and teammate routing. Internal comments support @mentions to notify specific teammates.
• Both connect via API key. Find them under Connectors → Support with built-in agent templates to get started.

Welcome conversation after onboarding

• New orgs that finish onboarding with at least one connector now land in a welcome chat instead of a static "you're all set" screen. Macha uses the connected tools to pull a few real items (recent tickets, recent payments, recent conversations) and the assistant message streams in referencing what was actually found.
• Three AI-generated suggestion chips appear below the welcome message, tailored to the agents you picked and the data Macha found.
• The whole conversation is on us, a "This chat costs no credits" badge sits next to the model picker. One welcome chat per organization.

June 17, 2026

Credit costs reduced for GPT-5 and GPT-5.4

• GPT-5 is now 2 credits per response (was 3). The most balanced model in the OpenAI family, capable enough for nuanced replies, cheap enough to run at volume.
• GPT-5.4 is now 3 credits per response (was 5). The highest-capability OpenAI model in the platform, with notably better instruction-following than GPT-5 for agents with strict "do this exactly once" rules and complex branching logic.
• GPT-5.4 Mini remains at 1 credit, still the default for most workflows.
• Same plan, more responses. Pro customers now get ~67% more GPT-5 responses and ~67% more GPT-5.4 responses per month within their existing 10,000-credit allowance.

Soft-delete agents, 30-day trash with restore

• Delete is now reversible. When you delete an agent, it moves to a Trash tab on the Agents page for 30 days. You can restore it any time within that window, all its triggers, instructions, sub-agent links, and tool config come back intact. After 30 days it's permanently deleted.
• New "Trash" tab on the Agents page shows trashed agents with how long ago they were trashed and how many days remain before automatic purge. One-click restore button per row.
• Delete confirmation copy updated to make the 30-day restore window explicit. No more "this cannot be undone", because now it can.
• Triggers deactivated, not deleted. Trashed agents' triggers are paused so they stop firing, but their webhook secrets and config are preserved. Restoring the agent brings the trigger config back; you re-enable them when ready.
• AI sidekick respects the trash flow, if you ask the dashboard AI to delete an agent for you, it now uses the same soft-delete path. Same 30-day restore window.

Model picker, clearer scoping

• Toast notification when you change the model from the top-right dropdown, clarifying that the selection only applies to the current conversation, and pointing you to Settings to change the org default if that's what you actually wanted.

Billing, top-up balance visible on Usage card

• Top-up credits chip now appears next to your monthly credit balance on the Billing page Usage card when you have a top-up balance. A green pill (+N top-up) sits above your monthly count, with a question-mark hover showing how top-up credits work (persistent across cycles, used automatically after monthly allowance is spent).

Reliability fixes

• Subscription recovery after card decline, if a Chargebee renewal payment fails and the retry succeeds, the org's subscription status now flips back to active automatically (was previously stuck in past_due, which caused webhooks to be rejected with 410 Gone). Affected customers' rejection counters now reset on recovery too.

June 10, 2026

Sub-agent runs are now visible in chat

• See exactly what each sub-agent did, Every sub-agent invocation now produces its own conversation record, linked back to the parent. In chat, a new card appears under each delegated message showing the sub-agent's avatar, handle, model, connector logos, and tool count, with a green "Handed off" badge. Click to expand and see the sub-agent's full nested conversation, system prompt, every tool call with arguments, every tool result, and the final reply, rendered with the same styling as the main chat.
• Lazy-loaded with no setup, Works automatically for every new sub-agent run. No configuration needed.
• Available for both chat and autonomous triggers, Whether the parent agent was called by a user in chat or fired by a webhook trigger, the sub-agent's run is captured the same way and shows the same expandable card.
• Sub-agent runs stay separate from your main history, They don't pollute the user-facing conversation counter, but stay available for 45 days under the existing TTL policy.

Zendesk tag updates now reliable

• Tag tool now actually adds the tag, Fixed a long-standing bug where zendesk_update_ticket_tags would report success but silently fail to apply the tag to the ticket. The connector now uses a portable pattern that works correctly across all Zendesk account configurations.
• Existing tags preserved, Adding a new tag no longer wipes the existing tag set. Removing a tag only removes that tag and leaves the rest intact. The tool result reports the verified final tag list so the agent sees the truth, not just its intent.

Agents listing, new columns

• Model column, The agents table now shows each agent's model alongside the provider icon (GPT-5, GPT-5.4 Mini, Claude, etc.) so you can scan model selection across your full agent set at a glance.
• Tools & Triggers in one column, Tool count and trigger count are now on a single line with stacked connector logos, separated by a middle-dot. Hover either chip to see the full list of names.
• Sub-agents column, New column shows how many sub-agents each agent has wired up. Hover to see the names of the linked agents.
• Tooltip readability fix, Tooltips on the tool / trigger / sub-agent chips no longer get cut off by the table header. They now open below the chip with full text visible.

Agent editor, save reliability

• Sub-agents stay linked after save, Fixed a bug where opening an agent in the builder and saving any change (even just the description) would silently clear its sub-agent links. Sub-agent assignments now persist correctly across edits.
• Agent details refresh correctly after save, After saving, the sub-agent list in the configuration panel now correctly shows each sub-agent's name, handle, avatar, and model, instead of showing a generic "Agent" placeholder. Same fix applies to the agent details page after restoring from version history.

June 5, 2026

Studies, Report view

• Analytics view per run, Every completed Study run now has a Report view (open it from the toolbar on the run page). Top-of-page stat callouts summarise the run; per-column aggregation cards visualise each reportable schema column.
• Bar charts for Yes/No, Single choice, and Multiple choice columns, Each bar shows the bucket label, count, and percentage. Click any bar to drill into the matching records in a modal that uses the same results table the main run page does, so navigating from a chart into the rows feels continuous.
• Histograms for Number columns, Min / mean / median / max summary stats at the top, then a histogram. With ≤10 unique numeric values each becomes its own bin; otherwise Macha bins automatically into ~10 equal-width ranges.
• Ask Sidekick preview bar, A wide horizontal input sits at the top of the report, shown as a disabled "Coming soon" preview. Study-scoped query tools land soon, once they do, you'll be able to ask the Sidekick free-form questions about a run's data and it'll answer using the run's results directly. Until then the input stays inert so it can't return misleading answers.
• Plan availability, Report view is Professional and Enterprise only. Trial and Starter plans see an upgrade card with a link to billing.

Studies, editor & results polish

• Tag-style options input, Single / Multiple choice column options are now added as pills: type each value and hit Enter (or paste a comma-separated list and Macha splits it). Click the × on a pill to remove.
• Inline column-type hints, A short info chip under the Answer type selector explains what each type returns, useful when you're picking the right shape for a question.
• Credits / record badge, Visible next to the Model selector in the Study editor and again at the top of every run page. The cost stops being a surprise.
• Run study confirmation modal, Clicking Run study now shows an explicit confirmation with record count, estimated credits, and the model. You confirm once before any work starts.
• Excel-like results table, Click any column header to sort (asc → desc → off). Search box in the toolbar filters loaded rows in real time. Sticky header with a clean divider so headers don't disappear as you scroll. Same table component now powers the report drilldown so navigating between views feels continuous.
• KB push: clearer expectations, The "Use as knowledge source" modal now shows how many records will become articles, your plan's source-limit usage, and a note that selected fields are merged into a single article body per record. Plan-limit-reached state is surfaced clearly before you commit.

Triggers, debounce on every event-driven trigger

• Debounce chip on every event-driven trigger, Previously only messaging triggers exposed a debounce setting. Now any Zendesk or Freshdesk ticket event, Slack mention or DM, or messaging event lets you click the Debounce chip and set a delay in seconds. Default 0 (off) for new event types; the existing 15s for messaging and 10s for Slack DMs are preserved unchanged.
• Cron and unkeyed Macha-webhook triggers stay free of the chip, Debouncing a clock-driven or per-org universal webhook doesn't make sense, so it's hidden by design on those triggers.
• Safeguards against runaway event streams, A debounce bucket is force-fired when it accumulates 25 events, or after a total wait of 10× the debounce window (capped at 2 minutes). Protects against malformed event sources holding the agent indefinitely.
• Better autonomous-conversation titles for custom webhooks, Zendesk and Freshdesk custom-webhook triggers now use the ticket_id and ticket_subject from the payload (when present) for the conversation title in agent history. Previously these showed up as "Custom Webhook" with no context.

Agent toggle, preserves your Zendesk Trigger config

• Soft-toggle on agent pause/resume, Toggling an agent off no longer deletes its Zendesk webhooks or Zendesk Triggers, it sets them to "inactive" on Zendesk's side instead. Toggling back on flips them to "active". Same Zendesk resource IDs throughout, so any Zendesk Trigger you built in your own Admin Center that references Macha's webhook keeps working untouched across pause/resume cycles. Explicit trigger deletes still tear down fully.

Connector-disabled tools, surfaced visibly on the agent builder

• Locked state in three places, When a connector disables a tool, every agent that had the tool enabled now shows it in a warning state on the flow pill summary (amber dashed with hover tooltip), the Add Tool modal (greyed-out row with a Manage button), and the enabled-tools card (amber banner with a Manage connector link).
• Manage connector in place, Each "Manage" link opens the connector's settings modal directly on the agent page. Re-enable the tool, save, and the agent's tools list updates immediately, no reload, no leaving the agent builder.
• Defense in depth, The backend already silently skipped connector-disabled tools at runtime. These UI cues exist so teams don't waste time toggling tools the platform won't actually run.

Billing fixes

• Top-up deep link scrolls correctly, "Need more? Buy a top-up pack" from the Studies page now scrolls to the top-up section after the billing data finishes loading. Previously the gated section didn't exist yet when the browser tried to scroll.
• Chargebee popup opens reliably on cold loads, Clicking Buy on a top-up pack now waits for chargebee.js to finish loading before opening the popup. Fixes a quiet failure where the first click on a freshly-loaded billing page silently did nothing.

May 30, 2026

Studies, batch AI extraction

• Run an extraction across thousands of records, Studies let you define a set of columns (the AI fills each one in) and run that extraction across an entire batch of records in parallel. The result is a structured grid: one row per record, one column per field, ready to sort, filter, export, or feed back to your agents.
• Zendesk Tickets as the first input source, Pick tickets with a Zendesk search query and optional date range. Choose which ticket fields the AI sees (subject, description, status, priority, tags, custom fields, full comment thread). More record sources are coming.
• Typed schema columns, Define each output column as boolean, single select, multi select, number, short text, or long text. Add per-column guidance so the model knows exactly how to answer.
• Pre-run estimate & review gate, Before any run, Macha shows you exact record count, credit cost per record, and total estimated credits. Nothing kicks off until you confirm.
• Test runs, Sample 10–50 records to validate your schema and guidance before committing to the full batch. Test runs are flagged in the UI so they don't clutter your real history.
• Cancel anytime, only pay for what ran, Cancellation is graceful. Credits are deducted per successful record, never up front. Errored and cancelled records cost nothing.
• 20,000-record cap per run, Hard ceiling to keep runs predictable. For larger studies, split by date range.
• Frozen run snapshots, Every run keeps a copy of the input config, schema, and model it ran with. Editing the parent Study later doesn't rewrite history.
• Push results to a Knowledge Source, Turn a completed run into a searchable knowledge source. Each result row becomes a document indexed via the existing embeddings pipeline, so the structured insights from your Study become searchable context for your agents.
• Plan availability, Available on Professional and Enterprise.

Credit top-up packs

• Buy credits on top of your monthly allowance, Three one-time packs available from the billing page: 1,000 credits for $99, 5,000 credits for $449, 10,000 credits for $799. Powered by Chargebee one-time checkout, completed in an in-page popup.
• Top-up credits never expire, Unlike monthly credits, top-ups roll over month to month. They're used automatically after your monthly allowance is spent, so you never have to think about which bucket to spend from.
• Available to paid subscribers, The top-up section appears for active Starter and Professional subscriptions. Trial users upgrade first; Enterprise plans get bypassed credit checks instead.
• Idempotent webhook handling, Top-up purchases are recorded with the Chargebee invoice ID as a unique key, so duplicate webhook deliveries can't double-credit an account.

Billing UI facelift

• Monthly & top-up balance shown separately, The Usage card now splits monthly plan usage (with a bar that resets each cycle) from your persistent top-up balance, with a note that top-ups are spent after your monthly allowance.
• "Buy a top-up pack" deep link from Studies, A "Need more? Buy a top-up pack" link sits next to the credits-remaining pill on the Studies Review screen. Clicking jumps straight to the top-up packs section on the billing page.
• Cleaner pack cards, Price, per-credit rate, and savings badge sit on a single row. The "Most popular" ribbon is now Macha purple to line up with the plan cards above.

May 2, 2026

Zendesk routing & ticket type

• Assign tickets to groups, The Assign Ticket tool now accepts a group, an agent, or both. Route incoming tickets straight to a queue (e.g. "Billing" or "Tier 2 Support") instead of always picking a specific person.
• List Groups tool, New read tool that lists every active Zendesk group, so agents can resolve a group name to its ID before routing.
• Update Ticket Type tool, New write tool to change a ticket's type, problem, incident, question, or task. Lets triage agents categorise tickets cleanly without touching custom fields.

April 23, 2026

New model lineup & pricing

• GPT-5.4 added, OpenAI's most capable model with 1M context window. Best-in-class reasoning and accuracy. 5 credits per message.
• GPT-5 is now the default, Powerful reasoning with 1M context. All new agents default to GPT-5. 3 credits per message.
• GPT-5.4 Mini updated, Now 1 credit per message (down from 1.25). Fast and affordable for high-volume agents.
• Retired models, GPT-5 Mini, GPT-4o Mini, and o4-mini have been removed. All existing agents on these models have been migrated to GPT-5.4 Mini.
• Updated pricing, Starter plan: $299/mo (3,000 credits). Professional plan: $699/mo (10,000 credits).

Connector health & state management

• Connector status tracking, Connectors now have a health status (connected, auth_failed). When authentication breaks, affected agents are automatically deactivated and admins are notified via email.
• Test connection, Verify connector credentials with a single click. Works for all connector types.
• Reconnect flow, API-key connectors show a reconnect form when auth fails. Credentials are validated before saving.
• Webhook rejection tracking, After a configurable threshold of rejected webhooks on broken connectors, returns 410 Gone to prompt external systems to disable the webhook.
• Refresh lock utility, Shared token refresh locking for OAuth connectors (Zendesk, Google, Shopify) prevents race conditions.

Agent builder improvements

• Model selector moved to Agent Details, Model selection is now in the agent configuration card with a helper notice explaining model differences.
• Expand editor, Full-screen modal for the instructions editor with Done/Discard actions.
• Add more buttons, Quick-add buttons on left flow cards for tools, triggers, sources, and sub-agents.
• Conversation history modal, Table view with search, date range, sort, infinite scroll, and agent column.

SSE streaming fix

• Real-time tool status, Fixed an issue where tool call progress ("Using Transcribe Audio") wasn't showing during execution. Switched from fetch ReadableStream to XMLHttpRequest for reliable SSE event delivery.

April 17, 2026

Image vision for Zendesk attachments

• AI reads image attachments, When an agent encounters an image attachment on a Zendesk ticket (JPEG, PNG, GIF, WebP), it downloads the image, encodes it, and sends it to the AI model for visual analysis. The agent can describe screenshots, read text in images, identify products, and parse receipts.
• Vision-capable models, Works with GPT-5.4 Mini, GPT-5 Mini, GPT-4o Mini, Claude Sonnet 4.5, and Claude Sonnet 4. Groq models gracefully return a "not supported" message.
• Shared image utility, Reusable download + base64 encoding helper available for all connectors. Future connectors (Freshdesk, Slack) can add image reading with a single function call.

GPT-5.4 Mini

• New model available, OpenAI's latest mini model with a 400K context window (3x larger than GPT-5 Mini). Stronger reasoning, better instruction following, same image vision support. 1.25 credits per message.

Zendesk custom fields tool

• Read custom field values with labels, New Get Ticket Custom Fields tool returns every custom field on a ticket with its title resolved from the field schema. No more raw field IDs, agents see "Bestellnummer: 013724827106" instead of "field_1900003089273: 013724827106".
• Schema caching, Field definitions are cached per connector for 5 minutes, so the schema fetch only happens once per connector window.

Zendesk subject update tool

• Update ticket subject lines, New tool to change a Zendesk ticket's subject. Validates 1-150 character length. Requires confirmation in chat mode.

Trigger context & conversation audit

• Structured trigger data, Each connector's trigger definitions now include a contextSchema that declares which payload fields to extract for the UI. Ticket ID, subject, and requester are stored on the conversation for frontend display.
• Raw payload preserved, The full webhook payload is saved on every autonomous conversation for audit and debugging.
• Richer conversation history, Autonomous conversations show connector icon + event label in the sidebar and agent history modal. Conversation titles use the extracted fields instead of the event label prefix.

Tool & trigger advisories

• Activation confirmation modals, High-impact tools (Add Public Reply, Update Ticket Status) and triggers (Every Customer Message, New Ticket) now show a confirmation modal when toggled on, with a severity-themed warning and custom confirm/cancel labels.
• Persistent advisory cards, Enabled tools and triggers display an advisory note on the agent configuration page, a short, always-visible summary of how to use the tool safely.

Zendesk OAuth improvements

• Refresh race fix, Concurrent refresh attempts no longer clobber each other's tokens. The race condition that caused the Schmid Digital outage (two webhooks racing on a single-process Node server) is resolved with compare-before-write logic.
• Reconnect flow, New Reconnect button in the connector modal for OAuth connectors. Re-authenticates in place, same connector ID, no tool stripping, no agent breakage. A new email template directs customers to the Reconnect flow.
• Re-auth preserves credentials, The OAuth re-auth path now preserves all existing credential fields (webhook IDs, signing secrets) instead of replacing the entire blob.

Credit alert improvements

• Persistent dedup, Credit usage alerts (50%, 80%, 90%) are now tracked in the database instead of an in-memory set. No more duplicate emails after deploys.
• Single admin notification, Admin credit alerts fire once per threshold crossing, not once per customer admin in the org.

UI improvements

• Tool activity results, Autonomous conversation tool cards now show success/error banners with the actual result (not just "Completed"). Failed calls show the error reason + dimmed args with "Attempted payload (not sent)".
• Image preview in tool cards, Image attachment results render a thumbnail preview with broken-image fallback.
• Scrollable org switcher, The organization dropdown scrolls when the user belongs to many orgs, keeping Settings and Log out visible.
• Past conversations expanded, The sidebar conversations list is expanded by default for paid plans.

April 9, 2026

Zendesk OAuth

• Global OAuth client approved, Zendesk has approved Macha's global OAuth client. OAuth is now the default authentication method for new Zendesk connections, replacing API key auth.
• Token refresh, Proactive and 401-fallback token refresh with per-subdomain locking. Tokens are refreshed automatically before expiry.

Confluence connector

• 7 tools, Search pages, read content, list spaces, get child pages, create pages, update pages, and add comments. Write operations require confirmation.
• Token expiry monitoring, Enter your Atlassian API token expiration date. Macha sends email reminders 30 days before expiry and automatically deactivates the connector on 401 errors.
• Auto-creates source, Connecting Confluence automatically creates a data source, just like Google and Notion.

AI Agent Builder & Editor

• Build agents through conversation, Describe what you need and the AI creates a fully configured agent with instructions, tools, triggers, sub-agents, and model selection. Available on all plans.
• Edit with AI, Open the AI editor from any agent's configuration page to modify any aspect of the agent through conversation.
• Artifact cards, Created and updated agents, triggers, and tools appear as interactive clickable cards in the chat.

Custom API Tools & AI Tool Builder

• Connect any REST API, Define HTTP endpoints with URL parameters, body templates, and authentication (API key, bearer token, basic auth). Credentials encrypted at rest.
• AI Tool Builder, Describe the API you want to connect, paste credentials, and the AI discovers endpoints, tests them, and creates tools automatically.
• Tool groups, Organize tools under named connectors with custom icons and brand colors. Groups appear on the connectors page.
• Built-in testing, Test endpoints directly from the tool editor before assigning to agents.

Scheduled triggers

• Cron-based agent execution, Schedule agents to run on recurring intervals (hourly, daily, or custom cron). 60-second polling loop with guardrails: min 1h interval, 24 runs/day cap, 3 org concurrency limit.
• Auto-disable on errors, Triggers disable after 5 consecutive failures with email notification. Stale lock cleanup at 30 minutes.

Sub-agents

• Multi-agent delegation, Agents can delegate tasks to specialist sub-agents, each with its own tools, model, and knowledge base. Up to 3 levels deep with circular reference prevention.
• Seamless UX, Sub-agent tool calls stream with indentation in the status log. Tools used by sub-agents appear in the "Tools Used" popup on saved messages.

Test Run

• Simulate with real data, Search for real tickets, emails, or documents from the agent config page. Select one and the agent processes it in a test conversation with live SSE streaming.
• Simulation schema, Every tool and trigger across all 14 connectors defines a simulation property for entity search and formatting.

Google Picker for document sources

• Native file selection, Select Google Docs and Sheets for agent data sources using the native Google Picker API instead of manually entering document IDs.

Custom webhooks (Macha trigger)

• Universal webhook URL, Generate a unique webhook URL with signing secret for any external service. No connector required, works with any system that sends HTTP webhooks.

File Tools connector

• Built-in file generation, Agents can generate XLSX spreadsheets, CSV files, PDF documents, and DOCX Word documents on the fly. Always available with no setup required.

GDPR compliance

• Scheduled deletion, 30-day scheduled deletion for accounts and organizations with cascade cleanup (Spaces files, MongoDB collections, org memberships).
• Data export, Download a complete copy of your data as JSON. Credentials, embeddings, and sensitive fields are stripped.
• Cookie consent, Cookie banner with accept/decline controls on all pages and layouts.
• Conversation TTL, Conversations auto-delete after 45 days of inactivity. Internal agent logs expire after 7 days.
• Credential encryption in logs, API keys and tokens in internal agent log messages are encrypted before storage using AES-256-GCM.

April 3, 2026

AI Agent Builder & Editor

• Build agents with AI, Describe what you want in plain language and the AI creates a fully configured agent with the right tools, instructions, triggers, and sub-agents. Available on all plans.
• Edit agents with AI, Open the AI editor from any agent's configuration page to modify instructions, add tools, set up triggers, or wire up sub-agents through conversation.
• Context-aware suggestions, The builder knows your connected integrations and suggests relevant agent configurations. Hover over suggestion pills to preview prompts.
• Artifact cards, Created agents, triggers, and tools appear as interactive cards in the chat. Click an agent card to navigate directly to its configuration.

Sub-agent enhancements

• Write permissions in autonomous mode, Sub-agents can now execute write operations (replies, updates) when running via triggers. They remain read-only during interactive chat for safety.
• AI builder support, The AI Agent Builder can assign sub-agents when creating new agents, with automatic circular dependency checks.

Agent configuration improvements

• Save guard, The save button is grayed out when there are no changes. A confirmation modal appears when navigating away with unsaved edits.
• camelCase handles, Agent handles are now camelCase (e.g., @ticketTriageAgent) and auto-generate from the agent name as you type.
• Inline delete, Delete agents directly from the agents table without opening the configuration page.
• Data source scope toggle, Inline "All documents" / "Selected" toggle on each data source row. No more expanding to switch scope.

April 2, 2026

Organization switcher

• Plan badges, Each organization in the switcher now shows its plan (Trial, Professional, Enterprise) as a compact tag, including trial expiry status.
• Full refresh on switch, Switching organizations now fully reloads the page to ensure all data reflects the new context.

Onboarding updates

• Show all connectors, A "Show all connectors" button on the connect step reveals additional integrations beyond the ones selected in step 1.
• Tools in connector modal, The connector setup modal during onboarding now shows the full tools list, matching the main connectors page.
• Zendesk simplified, Zendesk connection during onboarding uses API key auth only for a simpler setup experience.

UI refinements

• Template filters, Agent templates page now uses the same segmented filter tabs as the agents page, replacing the old pill-style filters.
• Sidebar refresh, The sidebar agent list now updates immediately when agents are created, deleted, or saved.
• Website crawler, Increased the page size limit from 512KB to 2MB, fixing ingestion failures on content-heavy sites like e-commerce stores.

March 30, 2026

AI Tool Builder

• Build custom API tools with AI, Describe the API you want to connect, paste your credentials, and the AI discovers endpoints, tests them, and creates tools automatically. Supports any REST API with API key, bearer token, or basic auth.
• Auto-grouping, Tools are automatically grouped under a named connector with emoji icon and brand color.
• Smart auth handling, Credentials are captured from successful test requests and applied to all tools in the session.
• Auto-detect tool type, POST/PUT/PATCH/DELETE methods are automatically marked as write operations requiring confirmation.

Custom Tools improvements

• Masked credentials, API keys now show a masked hint in the edit modal so you can verify which key is configured without exposing the full secret.
• Inline delete, Delete tools directly from the tools list without opening the edit modal.
• Better parameter handling, Empty optional parameters are automatically stripped from request URLs, preventing errors from APIs that reject empty query values.

Privacy & data management

• Privacy settings page, New section under Settings to download your data, manage account deletion, and access all policies.
• Data export, Download a complete copy of your data (conversations, agents, connectors, tools) in JSON format.
• Cookie consent, Cookie banner with accept/decline controls on all pages.
• Automatic data cleanup, Conversations auto-delete after 45 days of inactivity. Chat attachments expire automatically.

March 22, 2026

Shopify connector

• Full Shopify integration, Connect your Shopify store via OAuth. Search products, look up orders, find customers, view discounts, and process refunds, all from your AI agent.
• 7 tools, search_products, get_order, lookup_customer, get_customer_orders, search_orders, get_discounts, and create_refund (with confirmation).
• Demo account, Pre-configured Shopify support agent available for testing.

Scheduled triggers

• Cron-based agent runs, Schedule agents to run automatically on a recurring interval. Minimum 1-hour intervals with daily caps and concurrency limits.
• Auto-disable on failure, Triggers automatically disable after 5 consecutive errors to prevent runaway executions.
• Plan limits, Pro plans get 3 scheduled triggers, Enterprise gets 20.

March 15, 2026

Freshdesk connector

• 15 tools for Freshdesk, Full ticket management including get/search tickets, add replies and notes, update priority/status/tags/fields, assign tickets, search contacts, and browse knowledge base articles.
• Attachment support, Read and extract text from PDF, DOCX, XLSX, CSV, and TXT attachments directly from Freshdesk tickets.
• Webhook triggers, Automate agent responses on ticket creation and updates via custom webhooks.

File tools

• Built-in file generation, Agents can now create spreadsheets (.xlsx), CSVs, PDFs, and Word documents (.docx) on the fly. Files are generated in-memory and returned as download links.
• Always available, No connection or setup required. File tools are built into every workspace.

March 8, 2026

Multi-instance connectors

• Connect multiple accounts, Link multiple instances of the same connector (e.g., two Zendesk accounts or three Stripe accounts). Each instance gets its own credentials and name.
• Automatic disambiguation, When multiple instances of the same connector type exist, tool names are auto-suffixed and descriptions tagged with the account name so the AI knows which one to use.

Sub-agents

• Agent delegation, Agents can now call other agents as tools, enabling hierarchical task delegation. Assign specialist sub-agents for specific workflows.

Onboarding

• Guided setup, New 3-step onboarding flow: select your tools, pick from recommended agent templates, and connect your integrations. Get up and running in under 2 minutes.

March 1, 2026

Trigger system

• Webhook automation, Agents can now run autonomously in response to external events. Supported triggers include Zendesk ticket events, Freshdesk updates, Slack mentions, and custom webhooks.
• Debounce, Configurable debounce intervals to batch rapid events before firing the agent.
• Custom webhooks, Generate a webhook URL with signing secret for any external service. Works with Zendesk, Freshdesk, or any system that can send HTTP webhooks.

Chat attachments

• Drag-and-drop file uploads, Attach images, documents (PDF, DOCX, CSV, XLSX), and audio files to chat messages. Drag and drop onto the input or use the file picker.
• Document extraction, Text is automatically extracted from uploaded documents and included in the conversation context.
• Image understanding, Uploaded images are sent as vision content for models that support it.

February 22, 2026

Agent data sources

• Knowledge base, Upload PDFs, spreadsheets, and documents as knowledge sources. Small files are injected directly into the prompt; larger files are chunked, embedded, and searched on demand.
• Live connector sources, Connect Google Docs and Notion pages as live data sources. Agents read the latest content in real-time, no re-uploading when documents change.
• Scope filtering, Choose between all documents from a source or hand-pick specific ones for each agent.

Billing

• Credit system, Per-model pricing with credit-based billing. Credits deducted once per complete assistant response. Enterprise plans bypass all credit checks.
• Usage alerts, Automatic email notifications when credit usage hits 50%, 80%, and 90% thresholds.
• In-app checkout, Subscribe and manage your plan directly within the app via Chargebee-powered checkout.