The MCP server is in early access. The interface may change.
Teach your agent
Already in Claude Code, Codex, Cursor, or another MCP client? Copy this prompt and paste it in — your agent will walk you through connecting Adapt and running your first query. Prefer to set it up by hand? The full steps are below.What you can do
Once connected, you can address Adapt with@adapt from your client and put the same agent that runs in Slack and the web app to work — now alongside whatever you’re already doing in your editor or terminal. A few examples:
Use your integrations
Endpoint
Authentication
Authenticate with your Adapt session token, sent as a Bearer credential:sessionToken cookie (Application > Cookies > app.adapt.com).
Multiple organizations
If your account belongs to more than one organization, also send anX-Adapt-Org header with the organization ID you want to use:
400 because it can’t tell which organization to act in. A 401 means the token itself is missing or expired.
Setup
Claude Code
Codex
Codex has nomcp add command that can set request headers, so configure the server directly in ~/.codex/config.toml. Create the file if it doesn’t exist and add:
X-Adapt-Org header alongside it:
config.toml at launch, so start or restart Codex to pick up the server. Verify it registered:
transport: streamable_http, the URL, and http_headers: Authorization=***** (masked). Inside a Codex session, /mcp lists the server and its ask_adapt tool.
This writes your token to config.toml in plaintext. It’s the 7-day token, so treat the file as a credential and refresh it when the token expires.
The ask_adapt tool
The server exposes one tool, ask_adapt, which runs an Adapt agent in your organization.
Three ways to call it:
- Start a conversation — pass
messagewith nochat_id. Adapt creates a new chat, runs the agent, and returns the reply. - Continue a conversation — pass both
messageandchat_idto keep going in the same chat. - Read a conversation — pass only
chat_idto return the full transcript.
chat_id: <id>. Save that value to continue or revisit the conversation later.
Agent runs are asynchronous and can take some time on longer tasks. The tool waits for the reply and returns it when the run finishes.