Changelog

2026-10-14 — v1.0.0Copy Anchor Link

• General availability — Harnix moves from beta to new production infrastructure: the database is backed up daily, with a copy kept in a separate account, and can be restored to any point in the last 14 days. app.harnix.vn, api.harnix.vn and the widget embed code are unchanged.
• Knowledge search — a new multilingual embedding model gives better knowledge_search results for Vietnamese and English documents and for questions asked in the other language. See Knowledge bases.
• Beta data — organizations, agents, knowledge bases and conversations from the beta were not carried over; create your organization again and re-upload your documents.
• Security — the console and sign-in pages were updated to fix published critical vulnerabilities.

2026-10-12Copy Anchor Link

• Docs — new page Build an MCP server for teams writing their own MCP server: transport, the four authentication methods (headers, OAuth authorization code + PKCE, client credentials, none) with the exact standards, parameters and responses Harnix uses, tool and annotation design, result limits, and an error-code table with server-side fixes.

2026-10-11Copy Anchor Link

• Widget — question cards (ask_user) appear in conversations: 1–4 questions, radio or checkbox options, an "Other" field for free text, and a 10-minute countdown. Tapping sends single-select questions immediately (tap-to-send); the Send button enables once all questions have answers. The message composer stays unlocked — typing in the chat composer also answers the question and resumes the same run. The card restores automatically on page reload; timed out questions collapse into an expired line. Seven new customizable keys in strings. See Widget.
• Console — the agent Tools tab adds a Built-in tools section showing platform tools (ask_user) with dedicated enable toggles. The run status filter adds Waiting for input (waiting_input). Traces display two QUESTION steps (asked: questions, modes, options; and answer: selections, free text, or expired, with who answered). See Tools and Runs & traces.

2026-10-10Copy Anchor Link

• Built-in tools (ask_user) — agents can ask the end-user 1–4 multiple-choice questions when a decision is needed to proceed; users can pick options or reply in their own words. Runs pause in waiting_input without holding a worker, waiting up to 10 minutes. The stream gains a question frame, conversations gain pending_question, and a new route POST /v2/runs/{run_id}/answer accepts structured answers. Sending a regular message while a question is pending answers it with free text and resumes the same run. The built-in tools catalog is enabled by default for new agents and managed in the console. See Tools, Streaming, Conversations, and Idempotency & cancel.

2026-10-09Copy Anchor Link

• Connectors — a write action through a connector (a tool the server doesn't mark readOnlyHint: true) now waits for the end-user to confirm it in the conversation before it runs, showing the tool's name and real arguments; declining, or not answering within 5 minutes, makes the assistant say so and the conversation carries on. The stream gains an approval frame, conversations gain pending_approval, a new route POST /v2/runs/{run_id}/approval approves or declines (optionally for the whole conversation), and two new error codes 409 approval_pending and 409 run_not_waiting bring the total to 21. See Connectors, Streaming and Idempotency & cancel.
• Widget — a tool confirmation card appears in conversations: connector name, tool name, real arguments, three buttons (Approve, Approve for this conversation, Decline), and a 5-minute countdown. The message input is locked while waiting; reloading the page restores the card. Ten new strings override labels on the card, buttons, and the decision line. See Widget.
• Console — the Tools tab on an agent: each tool has its own Needs approval switch, annotation chips (Read-only / Write / Not declared), and a Reset to default button when overridden. The status badge chips add Waiting for approval (warn tone); the status filter has this value. Traces show two APPROVAL steps (asking for confirmation with tool details, and the decision with who/when/outcome). Conversations show a decision line under the assistant's message. See Connectors, Runs & traces, and Conversations.

2026-10-04Copy Anchor Link

• Console — the admin console now has Vietnamese for every organization screen (Overview, Agents, Knowledge bases, Runs, Conversations, End-users, Usage, Connectors, API keys, Integration, Members, Quotas, Organization). Switch between VI | EN at the bottom of the left sidebar, or at the top right of the sign-in page; the choice is saved to your account and the sign-in pages follow it. New accounts default to Vietnamese. See Overview.

2026-10-02Copy Anchor Link

• Limits — your organization now has a cap on chat runs at once (10 by default) and on chats waiting for a slot (50 by default), set by the Harnix team. A message that would only join a full queue gets the new error 429 queue_full (with retry_after and Retry-After) instead of waiting indefinitely; there are now 19 error codes. See Limits and Errors.
• Usage & quotas — Concurrent runs and Queue depth show your organization's caps instead of "no limit set", and count chat runs only.
• Widget — on queue_full the widget shows "Assistant is busy, try again in N s" and re-enables input when the countdown ends (override it with the new queueFull string). When a daily token budget is reached the widget says "Daily limit reached, please try again later" (new budgetExceeded string) and keeps the composer locked instead of offering a Retry that would fail. See Widget.
• Knowledge bases — up to 2 documents are processed at a time per organization (shared with other background work); chat is scheduled ahead of document processing.

2026-10-01Copy Anchor Link

• Runs & traces — a finished run's trace has its step-by-step content (prompts sent to the model, tool calls and their results) removed once it's older than your organization's retention period (30 days by default, set by the operator). Status, duration, iteration count, token totals, the run's input and its final answer stay; a purged trace shows Content not retained. Conversations and usage figures are not affected. Usage & quotas now shows Retention (a number of days, or Kept forever). See Runs & traces.
• Knowledge bases — a document's row menu now has Retry (for a failed document) and Re-ingest (for a completed one): both re-process it from the original file, no re-upload needed, and the list updates on its own. See Knowledge bases.
• Connectors — turn a connector off and back on with Deactivate / Activate in its menu. A connector whose tools haven't been listed yet shows — instead of 0 tools. Dialogs taller than your screen now scroll, so long tool lists in Add connector are fully reachable on a laptop.

2026-09-30Copy Anchor Link

• Connectors — add remote MCP servers yourself in the console under Connectors: sign in with OAuth in a popup, or use client credentials, a header, or no authentication; see each server's tools; choose per agent, on its Tools tab, which tools it may use (all on by default). A revoked sign-in shows as Needs reconnect with a one-click Reconnect. See Connectors.
• Docs — this documentation site, in Vietnamese and English, at harnix.vn/docs.
• Widget — the launcher is now a toggle: it stays visible while the chat is open, shows a close icon, and closes the chat when clicked again. The header close button works again, and the launcher icon no longer gets squashed.
• Admin console — the Revoke button on API keys stays inside its table; info boxes keep their text vertically centred; the knowledge base page no longer shows an outdated notice saying uploads are not processed. Each agent's page now has a real Usage tab (today's budget, then the usage report narrowed to that agent) in place of the earlier placeholder.
• Sign-in — the browser tab reads "Harnix", and the "Complete your profile" page (asked once when your account has no name yet) is styled like the rest of sign-in.

v0.1.0-beta — 2026-09-29Copy Anchor Link

The first beta release, live in production.

• Agents — provider/model configuration (Ollama, OpenAI, and any OpenAI-compatible endpoint), system prompt, history window, activate/deactivate, per-agent daily token budget.
• Knowledge bases — upload PDF/DOCX/TXT/Markdown documents (up to 20 MB), automatic semantic search (knowledge_search) once an agent has a knowledge base, answers with source citations.
• MCP servers — register an external server for an agent to call tools on, headers encrypted at rest, connection testing, per-agent assignment.
• Widget <harnix-chat> — embed with a one-line script, bubble or inline mode, light/dark theme, customizable color and labels (including Vietnamese translations).
• Integration — anonymous or signed token exchange, calling /v2/chat directly from your backend, streaming answers over SSE, cancelling a running run, conversation history.
• Admin console — Overview, step-by-step Runs & traces, Conversations & end-users (block/unblock/delete), Usage & quotas, API keys, Members & roles, the Integration page.
• Limits & isolation — rate limits per end-user and per API key, daily token budgets for the organization and per agent, full data isolation between organizations and between end-users.