Toolcaise MCP server

Connect a client

The server is hosted by Toolcaise at https://toolcaise.com/api/mcp (Streamable HTTP). Clients that support MCP sign-in open your browser: you sign in to Toolcaise, pick the workspace and the access, and approve. There is nothing to install and no key to copy.

Claude Code

Add the server, then run /mcp in Claude Code and choose Authenticate.
Sign in to Toolcaise, pick the workspace and approve.

Terminal

claude mcp add --transport http toolcaise https://toolcaise.com/api/mcp

Codex

Add the server, then log in to it.
Sign in to Toolcaise, pick the workspace and approve.

Terminal

codex mcp add toolcaise --url https://toolcaise.com/api/mcp
codex mcp login toolcaise

ChatGPT

In ChatGPT, open Settings, then Apps & Connectors, and turn on Developer mode under Advanced settings.
Create a connector named Toolcaise with the MCP server URL https://toolcaise.com/api/mcp and OAuth authentication.
Sign in to Toolcaise when asked, pick the workspace and approve.

Claude (web and desktop)

In Claude, open Settings, then Connectors, and choose Add custom connector.
Name it Toolcaise and use the URL https://toolcaise.com/api/mcp.
Choose Connect, sign in to Toolcaise, pick the workspace and approve.

Cursor

Use the install link, or add the server to ~/.cursor/mcp.json yourself.

Add to Cursor

~/.cursor/mcp.json

{
  "mcpServers": {
    "toolcaise": {
      "url": "https://toolcaise.com/api/mcp"
    }
  }
}

VS Code

Use the install link, or run the command below.

Add to VS Code

Terminal

code --add-mcp '{"name":"toolcaise","type":"http","url":"https://toolcaise.com/api/mcp"}'

Any other client, CI or scripts

Create a token below, keep it in an environment variable, and send it as a bearer token.

Streamable HTTP

URL:    https://toolcaise.com/api/mcp
Header: Authorization: Bearer $TOOLCAISE_MCP_TOKEN

Client menus change between versions. If a step above no longer matches, look for “MCP”, “connectors” or “custom server” in the client's settings and use the URL above.

Wire an agent with your coding agent

In Claude Code, Codex, Cursor or any coding agent connected to Toolcaise, open the repository that contains your agent and ask it to connect the agent to Toolcaise. The server gives it what it needs: the wire_agent prompt lays out the steps, and it finds the entry point, creates the agent with create_agent, stores the credential in the agent's environment, adds the SDK or exporter, and confirms the first report with check_agent_connection.

Ask your coding agent

Connect the agent in this repo to Toolcaise and check it reports.

It picks one of these paths, or you can name one:

Path	What it gives you
`python`	Python SDK: runs, spans, outcomes, model-call instrumentation, checkpoints and approvals.
`node`	Node.js SDK: the same as the Python SDK, for Node 20+.
`opentelemetry`	OpenTelemetry: send the traces your framework already emits. Observes only.
`connect-run`	Toolcaise Connect supervised run: wrap a command to meter model calls on the machine and stop it at a cost, token, call or runtime limit.
`connect-mcp`	Toolcaise Connect MCP gateway: put Connect in front of an MCP server to allow, deny, budget or hold tool calls for approval.
`coding-agents`	Claude Code and Codex: Connect turns on their own telemetry for the team. Observes only.
`http`	HTTP API: any runtime that can make an HTTPS request.
`langchain`	LangChain and LangGraph, through OpenInference and OpenTelemetry.
`crewai`	CrewAI, through OpenInference and OpenTelemetry.
`openai-agents`	OpenAI Agents SDK, with the Python SDK and an instrumented OpenAI client.
`claude-agent-sdk`	Claude Agent SDK, with the Python SDK and Claude Agent SDK hooks.
`vercel-ai-sdk`	Vercel AI SDK 7, through @ai-sdk/otel and OpenTelemetry.

In ChatGPT or Claude on the web, which cannot edit your code, ask for the setup steps instead: they come from the same guides as these docs.

Tools

Tool	Changes anything	What it does
`get_workspace`	No	The workspace, your role and access, the plan, usage and what needs attention.
`list_agents`	No	Every agent with its health, last heartbeat, host, schedule and the controls it accepts.
`get_agent`	No	One agent: credential, limits policy, recent runs and open incidents.
`list_runs`	No	Recent runs, filtered by agent or status.
`get_run`	No	A run's spans (model and tool calls, tokens, cost), outcomes and approvals.
`list_incidents`	No	Missed schedules, stale heartbeats, stalls and repeated failures.
`list_approvals`	No	Approval requests and their decisions.
`list_outcomes`	No	What runs produced, and whether a person accepted or rejected it.
`check_agent_connection`	No	Whether a newly wired agent has reported yet.
`get_setup_guide`	No	Exact steps to connect an agent by SDK, OpenTelemetry, a framework, Connect or HTTP.
`create_agent`	Yes	Registers an agent and returns its credential (once) with setup steps.
`rotate_agent_credential`	Yes	Issues a new credential and stops the old one.
`revoke_agent_credential`	Yes	Stops an agent's credential working.
`decide_approval`	Yes	Approves or denies one pending action.
`control_agent`	Yes	Pause, resume or cancel, applied at the agent's checkpoints or by Connect.
`set_agent_limits`	Yes	Sets or removes the limits Connect enforces for an agent.
`update_incident`	Yes	Acknowledges or resolves an incident.
`review_outcome`	Yes	Records a person's accept or reject decision.

Two prompts are also available: wire_agent and attention_review (what needs a person now, most urgent first).

Sign-in, access and tokens

A connection acts as the person who approved it. On every request Toolcaise checks that person's current role, so removing someone from the workspace, or demoting them, takes effect on their next call. Changing anything needs write access and an owner or admin role, the same rule as the dashboard; members get read access whatever they choose.

Signed-in clients get an access token that lasts an hour and a refresh token that lasts 30 days and is replaced each time it is used (OAuth 2.1 with PKCE and dynamic client registration). For CI, scripts and clients without MCP sign-in, create a personal access token in Dashboard, AI tools, and send it as Authorization: Bearer.

Everything a connection changes is recorded in the workspace audit log with the client's name. Disconnect a client or revoke a token in Dashboard, AI tools.

Agent credentials returned by create_agent are shown once, to the client that asked. Store them in the agent's environment or secret store, never in code.

What it does not do

It manages your Toolcaise workspace. It does not run your agents, and it cannot stop an agent any more than the dashboard can: SDK pause and cancel are cooperative, and only a run under Toolcaise Connect is stopped at a limit.
It cannot tell whether an agent's work is correct. The tools that approve actions or review outcomes record a person's decision; the assistant should only use them when you have told it what to decide.
It does not see prompts, completions or the arguments of held MCP tool calls; Toolcaise never receives them.