Connection tools
What the connector gives the model in Buff Code: trackers, documents, the web, your own MCP servers; why writing tools are included; where calls show.
Everything connected to your company’s connector is available to the model in Buff Code: task trackers, knowledge sources — Confluence, Notion, Google Drive, folders — the web and your own MCP servers. Nothing extra needs to be set up: the model gets the tools of the connector the developer signed in through.
What it looks like in Buff Code
Connection tools arrive as a single MCP server, buff (in the interface — /mcps). A tool’s
name is made of the connection and the action: handbook__search,
jira__get_issue. Next to it are two more servers from the same gateway: knowledge
(knowledge base) and repos
(project repositories).
The notes for the AI that the administrator wrote for connections reach the model together with the tools — just as in the platform’s work.
Writing tools are available
The developer works under their own name, so the model in Buff Code can not only read but also act: leave a comment on an issue, create a ticket, add a note to the handbook — with whatever permissions the connection’s account has.
This differs from the platform’s work. There, for your own MCP servers, the model gets only the tools the administrator marked as read-only, and in “Ask” and research the connections work read-only. In Buff Code the “read-only” mark reaches the model as a hint, not a prohibition.
What follows from this:
- The connection’s permissions = the model’s permissions. If a connection does not need to write, give it a read-only account — this is the only reliable limiter.
- Buff Code asks for permission to call a tool, just as for editing files — the permission mode is set in Buff Code itself.
- Every call is recorded twice: in the development session — with the tool’s arguments and answer — and in the connector’s call log with the developer’s name.
Where every call is visible
| Where | What is visible |
|---|---|
| The development session in the workspace | the call, arguments, the tool’s answer, the time |
| The connector’s local page, the “Buff Code” section | per developer: how many calls, how many with an error, the last call |
| The connector’s call log | every call marked “developer”, with the developer’s name |
What the model does not get
- Connection secrets: tracker tokens and MCP server keys stay on the connector.
- Connections of the organisation’s other connectors: only those of the one used to sign in.
- Platform tools that connections do not have: Buff tasks, reviews and audits are not managed from Buff Code — the workspace is for that.
Own providers and local models
Your own provider in Buff Code: OpenRouter, Anthropic, OpenAI, LiteLLM, Ollama, LM Studio. What goes to Buff and what stays on the machine.
Knowledge base
How the model in Buff Code reads the organisation’s and project’s knowledge base, proposes a new entry, and why an entry appears only after you confirm it.