The Dosu MCP Server provides coding agents with direct access to your team's knowledge via the Model Context Protocol. Instead of guessing from training data, an agent can search your Documents and Topics, read notes other agents on your team have written, and pull in answers your team has already approved. This page covers what the server exposes, which clients work with it, how to connect, and the workflow that gets the best results.

What it is#

MCP (Model Context Protocol) is an open standard that lets AI agents call external tools. The Dosu MCP Server is a hosted MCP endpoint that exposes your organization's knowledge as a set of tools an agent can use mid-task. It is managed from the MCP servers page in the Dosu app, where each server is scoped to the Sources you select.

The same server also carries built-in knowledge from hundreds of public open source libraries (projects like Airflow, vLLM, Prisma, and Zod). That gives an agent grounded context about your dependencies, not just your own code.

Why it matters#

Coding agents are only as good as the context they are given. Run them at scale, and the cost adds up. Agents re-read the same files, miss decisions buried in past issues and chat, and produce inconsistent work because each one starts cold. The Dosu MCP Server provides every agent with a shared source of truth, so answers stay consistent across people, repos, and tools, and knowledge lost to turnover remains accessible.

Tools the server exposes#

The server provides a focused set of tools, grouped by what they do:

• Retrieve and write your team's knowledge with read_knowledge and write_knowledge.
• Review pending changes with review_knowledge.
• Summarize a turn with finalize_session_knowledge.
• Check the connection with whoami.

For what each tool takes, returns, and when to reach for it, see the full tool reference in Connecting MCP.

Supported clients#

The Dosu MCP Server works with any MCP-compatible client, including:

• Claude Code
• Codex
• Cursor
• VS Code Copilot

Each server in the MCP servers page shows copy-paste connection instructions for these clients.

How to connect#

The fastest path is the Dosu CLI, which configures your client for you.

1. Open the MCP servers page. A default server is created with your organization and linked to your Sources. To scope a server to specific Sources, create a new one and select them.

2. Run the setup command in your terminal:

   ⧉

   npx @dosu/cli setup
   

   If you prefer Homebrew, install the CLI with brew install dosu-ai/dosu/dosu, then run dosu setup. To target a specific server, append --deployment <your-deployment-id>.

3. Follow the prompts to pick your client and authenticate. Restart the client if it does not pick up the new server right away.

   During setup, if your organization has no GitHub repositories connected, you'll be prompted to connect one so the MCP Server can answer questions from your code. You can connect a repository at this point or skip and add Sources later.

   When configuring Codex (version 0.147+), the CLI automatically writes omit_tools_from = ["deferred"] to the Dosu server entry in both the global ~/.codex/config.toml (or $CODEX_HOME/config.toml) and project-specific .codex/config.toml files. This ensures Dosu tools are loaded up front and immediately available, rather than being deferred behind tool_search.

Manual configuration#

If you would rather configure a client by hand, the server uses HTTP transport. Use the path-based endpoint shown on the MCP servers page:

https://api.dosu.dev/v1/mcp/deployments/<your-deployment-id>

For authentication, OAuth is recommended where the client supports it, since it handles credential rotation for you. Clients without OAuth can authenticate with an API key passed in the X-Dosu-API-Key header. Create one in Settings, Account, API keys.

Recommended workflow#

An agent follows a read-first pattern:

1. read_knowledge to load curated context for the task, along with the notes for its working branch when it passes repo and branch.
2. write_knowledge when the work surfaces something durable that the next agent would otherwise have to rediscover.
3. finalize_session_knowledge once at the end of the turn, with the receipts from the first two steps.

A note from write_knowledge is available immediately or after the pull request merges, depending on how it is written. When write_knowledge receives only repo (no branch), the note is unanchored and promoted right away. When both repo and branch are supplied, the note is branch-scoped and waits for its pull request to merge into the default branch before becoming a candidate Topic. Either way, the note does not wait in Review. Review holds Document changes and draft replies, scoped to the connected Library and MCP deployment rather than the whole organization. Agents manage it through review_knowledge only when you ask. To review another Library's queue, connect the MCP client to that Library's MCP deployment.

Notes#

• A server only reaches the Sources selected for it, so scope servers to match what each agent should see.
• The Dosu CLI learner writes notes with the session's repo and branch, anchoring them to that branch. They are promoted when the branch's pull request merges into the default branch, or immediately if the session was on the default branch or its pull request has already merged. Sessions with no known repo or branch are still studied; their notes are saved unanchored and promoted immediately.

Related#

• Dosu CLI
• Sources
• Connecting MCP
• Documents and coding agents (FAQ)