Introducing Agents.·Read
OverviewAgentsWorkflowsExecutionDeveloper ToolsSecurityLibrary

MCP

The Trickest MCP server. Your agent runs the graph.

Hosted at api.trickest.io/mcp. Sign in once from Claude, Claude Code, or Codex, or paste an API token into Cursor. The agent then builds and runs workflows from its own loop, with 40 tools it can read before it calls.

Connect itMCP docs

One URL. Sign in, or paste a token.

No package and no local server. Clients that run OAuth open a Trickest sign-in and get their own connection, listed and revocable under Settings › Credentials. Clients that only take headers send a Settings API token with the Token scheme.

Sign in · Claude Code
claude mcp add --transport http trickest https://api.trickest.io/mcp# then /mcp in the session, approve in the browser

In Claude on the web or desktop: Settings › Connectors › Add custom connector, paste the URL, connect. Codex: codex mcp login trickest.

Token · mcp.json
{  "mcpServers": {    "trickest": {      "url": "https://api.trickest.io/mcp",      "headers": {        "Authorization": "Token <YOUR_TOKEN>"      }    }  }}

Cursor, CI, or any Streamable HTTP client. Token from Settings › Credentials. Setup per client.

A call from your agent

The client discovers Trickest tools and runs them from its own loop. Same MCP mark every compatible host shows.

Agent · MCP

Thought briefly

Thought for 1s

Fleet is warm. Running the chain serially via MCP, starting with gau-history.

Thought for 2s

Workflow work from the agent

Build, distribute, and read results. Same engine as the CLI, called as tools.

Read the graph

Inspect a workflow before changing it: the graph and how data moves.

› Show me the discovery workflow

Add scripts

Put Python or Bash beside library tools as first-class steps.

› Drop a Python step after httpx

Wire and nest

Link ports by schema, or embed a whole workflow as one node.

› Connect outputs into the next module

Distribute the run

Execute on managed or self-hosted machines. Parallel where the graph allows.

› Fan this scan across the fleet

Query results

Read Live Tables and declared outputs after the run lands.

› Which hosts are still open on 443?

Run and wait

Start a run from the agent loop and wait for status before the next step.

› Execute and tell me when it finishes

Every tool says what it does before it runs

Read, write, or destructive is on the tool itself, so Claude, Cursor, and Codex can allow reads and ask on the rest. You keep the final click on anything that changes your workspace.

Read
23 tools

Graphs, runs, outputs, tables, memory. Set to always allow and lookups run without a prompt.

› trickest_run_get

Write
10 tools

Add nodes, wire ports, create workflows, stop a run. Your client asks before each call.

› trickest_node_add

Destructive
7 tools

Delete, replace, or spend compute. Deletes also refuse to act until the agent passes confirm: true.

› trickest_workflow_delete

Revoke a signed-in client any time under Settings › Credentials. Every tool and its effect.

questions

Frequently asked

Need the full setup? Open MCP docs.

No. The Trickest MCP server is a hosted Streamable HTTP endpoint at api.trickest.io/mcp. There is no npm package and no local binary. Point your client at the URL and sign in, or pass a token. See MCP setup.

Any MCP-compatible client. Claude on the web and desktop, Claude Code, Cursor, and Codex are the documented paths, and any Streamable HTTP client can point at the same endpoint. See Client config.

Two ways. Clients that run OAuth, such as Claude, Claude Code, and Codex, open a Trickest sign-in page; you approve once and the connection shows up under Settings › Credentials, where you can revoke it. Clients that only take headers, such as Cursor, send a Settings API token with the Token scheme. Read Sign-in and token setup.

Forty tools: search the library, read and edit workflow graphs, connect nodes, execute runs and wait for them, read outputs, verify results, query Live Tables, and manage variables and memory. Everything else in the CLI is reachable through a dispatch tool. The same run behavior as the CLI, typed for agents. See Tool reference.

MCP is for clients that already reason, like Cursor, Claude Code, and Codex. Use the SDK when you write the code and want a typed TrickestClient. Same engine, same token. See SDK overview.

A connected agent acts as you, under the same access controls as the platform. Every tool carries a read, write, or destructive mark, so your client can allow reads and ask before anything changes. Deletes also need an explicit confirm from the agent. Disconnect a client any time under Settings › Credentials. See How execution is isolated.

Same engine, other surfaces

MCP is one of four ways in. Pick the next cut of the stack.

Connect itMCP docs

Get a personalized demo

See Trickest in Action

A 30-minute walkthrough. We map the platform to your stack and answer pricing and deployment questions for your environment.

Book a demoExplore the platform