Skip to content
Tembo Agent Studio

MCP server

Tembo Agent Studio exposes a Model Context Protocol (MCP) server so an AI client like Claude Web, Claude Desktop, or Claude Code can read and improve your TAS deployment directly. The MCP server exposes live run history and output, the tool catalog, connection status, automations, Slack bots, and spec validation, plus the ability to trigger runs and hand authoring to the Tembo Coding Agent.

Connect Claude Web and Desktop

Section titled “Connect Claude Web and Desktop”

Claude Web and Claude Desktop use the same custom connectors. Add Tembo TAS once through Claude’s connector settings and it becomes available in both clients when you use the same Claude account or organization.

  1. In Claude Web, open the custom connector form:
    • Pro, Max, or Free: Customize → Connectors → Add custom connector.
    • Team or Enterprise owner: Organization settings → Connectors → Add → Custom → Web. Members connect it afterward under Customize → Connectors.
  2. Enter:
    • Name: Tembo TAS
    • Remote MCP server URL: https://<your-tas-host>/mcp
    • OAuth Client ID / Secret: leave both blank
    • Individual sign-in: keep enabled when the form shows it
  3. Click Add, then Connect. Claude opens TAS: sign in, choose the workspace this connector may access, review the requested access, and click Allow. No TAS API key or pre-registered OAuth client is required.
  4. In Claude Web or Desktop, enable Tembo TAS for a conversation and ask Claude to “list my agents” or “show the last failed run’s output.” There is no separate Desktop MCP configuration.

The TAS host must be publicly reachable over HTTPS, and its BETTER_AUTH_URL must exactly match that public origin. Each member completes their own OAuth flow, acts as their own TAS user, and chooses which workspace to expose.

Connect Claude Code

Section titled “Connect Claude Code”

The agent definition files live in your connected Git repo, which Claude Code can edit locally. Mint a key under Settings → API keys, then:

Terminal window
claude mcp add --transport http tas https://<your-tas-host>/mcp \
  --header "Authorization: Bearer tas_..."

That’s it — start a Claude Code session and ask it to “list my agents” or “show the last failed run’s output”. Any MCP client that speaks Streamable HTTP works the same way; point it at https://<your-tas-host>/mcp with the same Authorization: Bearer header.

Authentication

Section titled “Authentication”

Claude Web and Desktop use TAS’s OAuth 2.1 flow: Dynamic Client Registration, S256 PKCE, short-lived signed access tokens, and rotating refresh tokens. The consent flow binds the token to one workspace. Claude Code and header-based clients can use the same personal API keys as the REST API instead.

Both methods act as you: tools that run agents use your per-user connections, and every request resolves your live workspace role. Read tools need viewer. Write tools (trigger_run, create_automation, request_agent_change) need both the OAuth mcp:write scope (when using OAuth) and the operator role; a viewer cannot gain write access by authorizing a connector.

Tools

Section titled “Tools”

Read (viewer):

  • list_agents — every agent in the repo, including specs that fail to parse.
  • get_agent — one agent plus its raw spec text and any sidecar tools/skills.
  • validate_agent_spec — parse a draft spec without writing it. Use this before committing.
  • list_runs — recent runs, filterable by status / agent / trigger.
  • get_run — full output, error, and token usage for one run.
  • list_tools — your cached tool catalog. Each tool’s slug is what goes into an agent’s connections: tools: [...].
  • list_connections — your connection status across Composio + Native MCP.
  • list_automations — scheduled automations.
  • list_slack_apps — the workspace’s Slack bots (secret-safe).

Write (operator):

  • trigger_run — run an agent now, acting as you; returns a run id to poll with get_run.
  • create_automation — schedule an agent on a cron expression.
  • request_agent_change — hand an edit or a new-agent request to the Tembo Coding Agent, which opens a PR (or commits directly, per your commit mode).

Slack bots (workspace_admin):

  • create_slack_app — create a Slack bot (metadata only; comes up configuring and needs the one-time browser OAuth install before it’s live).
  • update_slack_app — change a bot’s name, the agent labels it may launch, owner, or secrets.
  • delete_slack_app — remove a bot.

A typical loop with Claude Code

Section titled “A typical loop with Claude Code”
  1. Ask Claude Code to read the agent you want to improve (get_agent) and the relevant runs (list_runs, get_run) to see what actually happened.
  2. It edits the agent’s YAML in your local repo checkout — its native strength.
  3. Before committing, it can validate_agent_spec to catch shape errors, and list_tools to get the exact connection slugs.
  4. After you commit and the change is live, it can trigger_run to test, then get_run to read the result — iterating until the agent behaves.

request_agent_change is there for clients that aren’t themselves coding agents (or when you’d rather TAS open the PR). When Claude Code can edit the files directly, that’s usually the faster path.

Notes

Section titled “Notes”
  • Stateless. The server runs in stateless mode — there’s no session to keep alive; each call is authenticated by the bearer token on its own.
  • Per-user connections. A run triggered through the MCP server uses the connections you authorized. If a run fails on a missing connection, check list_connections and authorize it under Connections.
  • The token is shown once. Revoke or disable a key anytime under Settings → API keys; the change takes effect immediately. OAuth clients manage their tokens automatically and can be disconnected from Claude’s connector settings.