Changelog for developers and admins

Newest first. Dates are DD-MM-YYYY. Servers get these changes with the next installer that includes them.

19-10-2026

Chess Coach, a new agent under Learning (live 18-10). Play the computer at Moderate or Difficult. Every move gets a mark from ★ Best to ?? Blunder and a plain note on why; mistakes show You played, Better was and the line. Three hints per game, takebacks, PGN export, and an accuracy summary at the end. Notes are built from the chess engine’s analysis.
Automation, a new agent under Research (live 16-10). Describe what to watch in plain words; it picks up to 5 of 30 news feeds and saves an alert for every new matching story, checked every 5 minutes on the server, with desktop alerts in Chrome while the page is open.
Alternatives answers in chat (live 16-10). Ask for alternatives to a product and get a sourced comparison table with checked prices, in rupees by default.
API & Usage page for administrators (live 17-10). Every API key with search, filters (Can add documents, Never used, Unused 30+ days) and the Add docs switch with Undo; the Top 10 by usage card moved to its Usage tab.
Website import (live 16 and 17-10). example.com and www.example.com count as one site; when the start page is refused the import tries the site’s sitemap; a pause asked for in robots.txt is shown and kept up to 30 seconds; an interrupted import ends as partial or failed instead of hanging. Import jobs carry crawl_delay_s, crawl_delay_source and crawl_delay_requested_s.
Chat attachment limits are now set in token-limits.cfg (live 18-10). Defaults are unchanged: 48,000 characters per file, 72,000 per message, 50 MB per file, 200 PDF pages.
AI Inbox: only your organisation’s mail and Other mail (IMAP) can be connected for now. Gmail, Yahoo Mail, iCloud Mail, Microsoft 365 and Google Workspace show OAuth policy to be configured (live 16-10). See AI Inbox.
Word files: asking for one Word document now makes one Word file instead of being refused as a word limit (live 17-10).
None of these are in installer 4.6.0; they come with the next installer.

17-10-2026

Add a document to a workspace through the API. POST /workspace/documents adds a .txt, .pdf, .docx, .xlsx or .csv file to one of the key owner’s own personal workspaces, and GET /workspace/documents/{id} returns its status (processing, not_ready, ready or failed) by your client_document_id or the server document_id. Every answer carries a limits object. Only keys an administrator ticks Can add documents may add, and the server must have it switched on. New codes include key_not_allowed, duplicate_name, id_in_use, file_too_large, file_too_complex, too_many_uploads and daily_upload_limit. See Add a document.
Copy-paste commands for Windows PowerShell and Linux or macOS, for listing, asking and adding, in the workspace API reference.
Not in installer 4.6.0; comes with the next installer.

15-10-2026

AI Inbox, a new agent under Email and notes. Connect your own mailbox (IMAP and SMTP, with a password or app password) and AI Inbox sorts mail into smart folders, writes an AI brief with amounts, dates and references, suggests next steps and drafts replies on the Smart tier. Every send is confirmed, with 5 seconds to undo. Mail stays on your server for 30 days; your mailbox is not changed.
Full screen for every agent: the button beside + New run, or Ctrl+Shift+F; Esc leaves it. See ZenithAI Agents.
Workspace API card and API guide tab. Each workspace’s settings show a Workspace API card with the workspace ID and a ready request, and Settings › API guide has a Workspace tab. See the workspace guide.
These three, and /workspace below, are not in installer 4.6.0 and come with the next installer.
New API: /workspace. Ask a workspace’s documents from your own software with an API key. GET /workspace lists the workspaces the key’s owner can open; POST /workspace answers from the whole workspace or one document, with found, sources, coverage, follow-ups through history, streaming and a plain text option. Answers never use the web and nothing is stored. Full reference: API reference: Workspace. Comes with the next installer after 4.6.0.

14-10-2026

Installer 4.6.0 for x86 servers (Rocky Linux 9 or later), the same build as the ARM64 installer. x86 servers now get everything below too. See Installation.
Installer 4.6.0 for ARM64 servers (Ubuntu 24.04 LTS). It has everything below, from 12-10-2026 on: /v1/responses, tools on Fast, functions that return images, ZenithAI Agents with CSV to Dashboard, Designer and Note Taker, and the new request limits. It also sets up Note Taker’s voice models.
Functions can return images on /v1/responses. A function_call_output now holds input_text and input_image parts, so an agent can fetch a scan, a photo or a chart from your systems and read it. Works with streaming, every tool_choice mode and JSON schema output. New example: Bills against the ledger.
Web answers read further. When the top pages of a web search can’t be read, the next results are tried before the answer is written.
Long pastes become a card in the chat box, so the message stays readable. The text is sent unchanged.
About Agents: the first switch from Chat to Agents shows a short note on what the agents are and how to build your own and connect it with ZenithAI. Tick “Don’t show this again” to hide it.

13-10-2026

Cleaner stream endings on /v1/responses: a stream that fails after text has started closes the open message as incomplete before response.failed, and leading blank deltas are no longer sent.
Designer in ZenithAI Agents: describe a page and get a finished offline .html file. When a request can’t be met in full, Designer still builds the page, with [placeholders] and a note of what was left out.
A collapsible left pane on the chat and Agents pages.
Web follow-ups in workspaces keep the subject of the conversation in the search.

12-10-2026

Function calling on Fast. Fast, Smart and Genius take tools on /v1/chat/completions and /v1/responses. Check capabilities.tools in /v1/models.
Request limits on /v1: 60 requests a minute per API key by default, plus per-address limits. A request over a limit gets 429 rate_limit_exceeded with Retry-After. See Errors and limits.
The OpenAI Responses API, POST /v1/responses: typed items, function calls, JSON schema output, typed stream events and the OpenAI Agents SDK. See the Responses guide.
Note Taker records meetings, names the speakers and writes minutes, with a Word download and a Recordings tab. See Note Taker.
A larger tool context on Genius, so longer tool conversations fit.

New on this site

API reference for Responses, Chat Completions, Models and Errors.
Three new agent examples: Bills against the ledger, Ask your database and Multi-agent help desk.
A downloadable example script with a function that returns an image.

Common questions

When do installed ZenithAI servers get the changes in this changelog?

Changes reach a server when it is updated with an installer that includes them. Pages that describe a newer feature say so near the top, and your administrator can tell you which build your server runs.