What’s new

October 2026

Connector v0.18.0: who may ask the bot, and an answer as private as the question (5 October). Every connection has a “Who may ask the bot” list (login, address, domain, regular expression) and a “Buff organisation members” checkbox; everyone else gets a polite reply from the bot that they can ask in Buff, and nothing is started. An answer to a GitLab internal note, a private Redmine note or a Jira comment restricted to a role or group is published just as privately. Redmine writes the answer and the link in the server’s markup (Markdown or Textile — a connection setting). The connection check and integration descriptions now use plain words. The connector page is written in plain language. In Buff Code the balance line and the “Balance and spending” window work. Integrations update from the catalogue; the connector — update it. See Who may ask the bot.

Buff Code 2.0.11-buff.16: a window and a console for everyday work with the agent (4 October). A sessions column with pinning, groups and an archive, search through the conversation; permission modes “Ask / Edit on its own / Full access” — an edit request shows the edit itself; a summary of changes under every turn and a safe rollback; a plan with approval, a to-do list and a /goal; a side conversation; a Git window with a commit message from the agent; scheduled automations; phased subagent scenarios; MCP servers, skills, commands, subagents, memory and hooks — from the settings; balance and spending in roubles (with connector v0.18.0); Russian and English interface. All of it is described in the new section Working in Buff Code.

Buff Code 2.0.11-buff.15 and connector v0.17.0: development sessions — recorded by Buff Code itself (2 October). The development session record is now kept by Buff Code itself: every message, piece of reasoning, tool call and its result reaches Buff as the work goes on, and the model server adds the model, tokens and charge to each turn. The “Development sessions” section of the web app reads this record: a call and its result — in one response, the JSON export — the session’s messages with response metrics; the “Call parameters” panel has been removed (the server no longer stores request bodies). Records made before the update have been moved to the new form without loss. Buff Code updates arrive by themselves; manually — buff upgrade; the connector — update it. See Development sessions.

September 2026

Buff Code 2.0.11-buff.10 and connector v0.16.0: the model under the plan’s name (28 September). If we chose the model for the organisation, Buff Code shows not its name but the plan’s name — in the line under the input field, next to answers and in the model picker. A model the organisation chose itself is shown under its own name, as before. Buff Code updates arrive by themselves; manually — buff upgrade. See Models and spending.

Confluence and Figma without a list mean everything open to the bot account (27 September). An empty list of Confluence spaces now means all spaces open to the bot account, not “nowhere” — as with Notion and Google Drive. Figma does the same for empty lists of files and projects: the bot reads any design by its link. Figma cannot list all of an account’s files, so without a list the bot learns about a question in a comment only from the FILE_COMMENT webhook — in the files it is set up on. The webhook comes from the Figma cloud, so the connector’s event receiver must be reachable from the internet; without the webhook such questions never reach the bot. Trackers, as before, answer nowhere without a list. Integration versions — confluence 1.0.6 and figma 1.0.4. See Knowledge sources.

Buff Code 2.0.11-buff.9: the /project command opens the project picker again (23 September). In earlier builds the command hint listed two identical /project entries, and Enter opened someone else’s session picker instead of the organisation’s projects. The update arrives by itself; manually — buff upgrade. See Commands.

Buff Code knows the knowledge base and the project’s repositories, and answers questions about sessions (23 September). The model in Buff Code reads the organisation’s and the project’s knowledge base and suggests new entries — they appear only after confirmation through a form in Buff Code. The project’s code in Buff is available to it without a clone on the machine: file search, content search and reading on any branch, read-only. The “Development sessions” section has a new “Questions” tab: ask the model about one session, about a selection or about a set of marked sessions. Requires connector v0.15.0 and Buff Code 2.0.11-buff.8. The Buff Code documentation has been rewritten: commands, troubleshooting, sessions on separate pages.

Buff Code — a development environment for those who write code themselves (22 September). A console and an app for Linux, macOS and Windows on your organisation’s models, knowledge and tools. Buff Code signs in through the company’s connector — with a code confirmed in the web app; the model is chosen by the administrator, spending goes on the organisation’s plan marked “Buff Code · device”, and all of the connector’s connections — trackers, documents, your own MCP servers — are available to the model as tools. Every model turn is recorded in full: a developer sees their own sessions, the organisation owner sees all of them, in the web app’s new “Development sessions” section, with export and an entry in the audit log. There is no separate plan. Requires connector v0.13.0 with the developer gateway. See Buff Code and “Download”.

Word, Excel, PowerPoint and PDF — wherever the model works (19 September). Documents can be attached to a task, to an answer to a clarification, to a review finding and to a question in “Ask” — and the model reads their content: text, tables, slides and their notes. Both generations of Office are supported (.docx/.xlsx/.pptx and .doc/.xls/.ppt) as well as PDF, with files up to 10 MB. The same documents are also read where they already live: in the project repository and in a space — both file reading and content search see them, so search no longer keeps quiet about what is written inside a document. What kind of file it is is decided by its content, not its extension. See Attachments.

Your own MCP servers in the connector (16 September). Besides integrations from the catalogue, the connector runs your MCP servers — from an archive, a GitHub repository or by address. Only the tools the administrator has marked as read-only are given to the model; without the administrator’s explicit consent the server does not start. Requires connector v0.12.0. See Your own MCP servers.

Notion — pages and databases as a knowledge source (16 September). The model reads Notion pages and databases when an answer needs them; questions are accepted in comments on a page, the answer arrives in the same place, and a research report appears as a child page. Requires connector v0.8.0. See Notion in the catalogue and its page.

Space — a workplace without code (16 September). A third kind next to development projects and repositories, for teams whose knowledge lives in documents rather than in code: product managers, analysts, support, customers. A space has “Ask”, deep research, the knowledge base and connectors — and no development, audits or review. “Ask” answers from the connected sources and the knowledge base with links to pages and entries; research compares documents and names contradictions. Created from just a name, instantly. The knowledge base and connectors now also open in an empty organisation. See A space without code.

Files on disk — local folders and SMB network shares as a knowledge source (16 September). The model reads the marked folders with the same tools as code: glob searches by name, grep — by content, read reads a file in full. Word documents, Excel spreadsheets and PowerPoint presentations — in both the 2007+ and 97–2003 formats — PDFs and text files are read as text; read-only, nothing is indexed, nothing is truncated. A network share is connected without mounting — the integration itself speaks SMB2 from the very first packet (SMB1 is not needed), signs in with NTLMv2 and signs the session. See “Files on disk” in the catalogue and its page.

Connector v0.11.0. Credentials of the password kind — the only kind the connector passes to the integration: for a protocol where the connector cannot insert the secret on the plugin’s behalf (SMB). The manifest declares this in advance, the catalogue shows “passed to the integration itself”, and every hand-over goes into the log. There is also a new network declaration — connections the integration opens itself (protocol, port, the field with the addresses); the connection page shows them next to the HTTP addresses. Integrations that need this require this version (host API 5).

Google Drive — documents, spreadsheets and presentations as a knowledge source (16 September). The model reads the marked Drive folders: a document as markdown, a spreadsheet sheet by sheet, a presentation as text, and searches content using Drive itself. Questions are accepted in comments on a document, spreadsheet or presentation, and the answer arrives in the same thread; a research report appears as a new document in the same folder. Two ways to sign in: a service account (folders are shared with its address) or your own account — sign in once through the browser, after which the connector refreshes access itself. See Google Drive in the catalogue and its page.

Connector v0.10.0. Two kinds of credentials that the connector issues itself: service_account (a JSON key; the connector signs a JWT and exchanges it for a token) and oauth2 (your own OAuth client; sign-in via a redirect to the connector’s local page or by pasting the redirect address, with the refresh token kept by the connector). An integration can offer several kinds — the sign-in method is chosen in the form and on the command line (--kind), and there is a new connector instance signin command. Integrations that need this require this version (host API 4). v0.10.1: signing in again replaces the access token immediately, not after an hour; the client file from the Google console is accepted as a whole. A question from a discussion now goes to the model together with the title, link and key of the place where it was asked — “what is this document about?” is answered about that document.

Connector: notes for the AI on every connection (16 September). A connection now has a free-text field for what the model should know about your system: which project does what, what you call things, where to look first. The text goes to the model together with the connection’s tools — as the tool server’s instructions and in the description of each tool. By default — up to 2048 characters; the limit is lifted with a checkbox after confirming the risk, and while it is lifted, the connection is flagged with a warning on the connector and in the app. Set on the connector page, on the command line (--ai-notes) and in YAML (ai_notes). See Tools and the model.

Review: changes by repository and file, without loading everything at once (12 September). The “Changes” tab opens with a list of files under each repository — with a mark of what happened to the file and line counters; a file’s content loads when you expand it, a long file in parts as you scroll, and its header stays pinned at the top. A review with a very large file (tens of thousands of lines) no longer fails with “result too large”; the “Patch” downloads at any size. On a phone the review screen scrolls as a whole. See Changes.

Knowledge base: scope-based rights, an honest conflict, reliable passes (9 September). The right to the knowledge base now works by the entry’s scope: a member with access to a project sees the entries of the project and the organisation and edits the project’s entries. The menu item is visible only to those with the right to read. On simultaneous editing, the editor shows the current text next to yours and lets you choose what to save. An entry’s address can be corrected by hand, the body has a preview, the author filter tells a person from the model, and the cost line expands into a list of entries by size. The model no longer rewrites your entries — it only marks them “needs review”, and the mark is cleared with the “Reviewed” button. Everything the model does with the base is recorded in the audit log with the stage and the task. A pass that did not finish is visible on the card and does not block the next one; the card shows what the pass cost. See Knowledge base.

The organisation’s knowledge base. A shared store of knowledge for people and the model: working rules, facts about the project and environment traps. The index of all entries goes into every model run, the bodies on demand, and entries marked “show to the model upfront” in full; above the list you can see what this costs in tokens. The model adds to the base itself after finished work — including a rejected or stopped task, where the lesson is most valuable — and then tidies the base up: merges duplicates, updates descriptions, removes its own outdated entries and flags discrepancies in your entries. There is no deletion: an entry goes to the archive with a reason and comes back with one click. Everything changes under permissions and is recorded in the audit log. See Knowledge base.

Single sign-on (SSO) for organisations. Employees sign in through the corporate provider — Keycloak, AD FS, Entra ID and any other over OIDC or SAML 2.0 — without a password or an email code. An account is created on first sign-in with the default role; existing accounts are linked by a verified email domain (a TXT record in DNS) or once with a password. “SSO only” mode closes password sign-in for members, while owners remain the fallback sign-in. Everything is set up by the owner in Settings → “Single sign-on”; every step goes into the audit log. See Single sign-on.

The organisation’s audit log. The organisation settings have a new “Audit log” section: who did what, when and from where — sign-in and its refusals, members’ rights, tasks, reviews, credentials, connectors, the plan, actions of the pipeline and of the platform operator. Records carry the connection’s address and browser, the result and the reason for a refusal; they are not edited or deleted, and each is linked to the previous one by a hash — integrity is checked with a button or with your own script on an export. Filters by period, actor, result and category; export to CSV and JSON Lines with no size limit. The owner reads it; a member can be given a separate right “View and export the audit log”. See Audit log.

The web for the model — from inside your perimeter. The “Web” integration gives the model search and reading of public pages through the connector: DuckDuckGo without a key, your own SearXNG, or Brave with a key. The limits are set by the connector administrator — an allow list or block list of sites with paths, content types, stop words; the internal network is always closed and cannot be opened by any setting. The model’s own page fetching is switched off everywhere: without this integration it has no internet. See “Web” in the catalogue and “The web for the model”.

Connector v0.7.0. A new kind of outbound rule — web: the address is named by the model, not the operator, so instead of a host the rule has lists, a check that the address is public with a connection made to exactly the checked address, and a ban on credentials on such a request. Integrations that need this require this version (host API 2); an older connector refuses to install them and names both versions. A setting with a fixed set of values is now picked from a list rather than typed in. v0.7.1: the connector’s reply channel works in a NATS queue — if two processes are running with the same identity (after a failed restart), a tool call is executed by exactly one of them, not both.

Figma — designs as a knowledge source. The model reads the marked Figma files the way the inspector panel shows them: screen structure, auto layout, colours in hex, typography, styles and variables by name, components and their properties — as parameters, without images, for any model. A rendered design is given only to a model that reads images: such a tool is marked “image” in the catalogue and is not offered to a text-only model. Questions are accepted right in comments on the design. See Figma in the catalogue and “Knowledge sources”.

Connector v0.6.0. An integration manifest can declare a tool that answers with an image (vision) and default setting values (default) — the form shows them straight away, and the command line fills them in itself. Integrations that need this require this version.

Tracker integrations — eight, from one catalogue. Jira Data Center and Server are joined by Jira Cloud, GitLab (self-hosted and gitlab.com), GitHub (including Enterprise Server), Gitea and Forgejo, Bitbucket Data Center and Bitbucket Cloud, Redmine. A question asked in a pull request discussion is worked out against the code of that branch; a request is written as a mention or as the tracker’s native slash command. See “Tracker integrations” and the integration catalogue.

The connector has become an integration host (v0.5.1). The core handles git; everything else is installed from a signed catalogue, runs as a separate process under a separate OS user and goes only to addresses derived from your own settings. What an integration will get is visible before installation.

The built-in Jira integration has been removed. Configured the old way (with a jira: section in the config), it no longer works — what to do.

Integrations run on the system Python. If no own runtime has been deployed, the connector takes Python 3.11+ from the machine — and says on its local page which interpreter it uses. An interpreter the connection’s user cannot reach (for example, in root’s home directory) is skipped with a clear explanation.

The command line now matches the form in credentials. connector instance secret takes the access kind from the integration manifest (a token, an “account + secret” pair, the tracker’s own header) and accepts --account. Previously a connection to Redmine or Jira Cloud set up by a script did not authenticate at all.

Confluence as a knowledge source. A new integration reads the marked spaces at the model’s request and accepts questions in comments on a page — one and the same for Data Center, Server and Cloud. Details are in “Knowledge sources”.

An address with a path finally works (v0.5.2). An integration configured with an address with a context path — https://company.atlassian.net/wiki, https://host/jira — was refused on every request: outbound rules are declared relative to the product and were compared with the full path. Now they are placed under the address the operator entered.

This section has been kept since September 2026; earlier changes are not reflected in it.