Skip to main content
Cargo speaks Model Context Protocol in two ways: The UUID path is Cargo’s address for a workspace resource — not an MCP protocol requirement. Linear and Notion use /mcp the same way the platform server does. Keep curated servers on the UUID path; OAuth tokens are audience-bound to that exact URL. Agent skills over the cargo-ai CLI stay the right surface for building: project deploys, schema changes, and large batches.

Connect the platform MCP

Hosted clients that speak HTTP (ChatGPT custom connectors, Claude.ai, Cursor over HTTP) connect to:
The client discovers Cargo’s authorization server, opens a browser for you to sign in, and then calls the server with a short-lived access token. If you belong to more than one workspace, the consent screen asks which one to bind. Workspace API tokens still work as Authorization: Bearer. On the MCP servers page, Connect Cargo shows this URL and the live OAuth grants against it.

What it can do

The platform server is a small, stable tool list. It does not dump every CLI command or every workspace tool into the client’s picker.
  • whoami / get_usage — signed-in user, workspace, and recent credit usage
  • search_actions — search connector, native, tool, and agent actions. Each result carries its price in credits[].cost. Without it, a connector or native action is free and a tool or agent is priced by the steps it runs; a tool with several entries costs between the lowest and the highest, depending on the path it takes. Cargo’s presets (such as the Find email and Find phone tools) are read-only tools and agents seeded into every workspace, and rank first when they match a query as well as a single provider’s action does
  • get_action_schema — input and output JSON schemas for those actions, without running anything
  • execute_action / execute_action_batch — run one input or a batch (credits apply)
  • get_run / get_batch / list_runs — poll a run or batch; get_run returns the last node output (not computed configs or traces); a finished batch carries a time-limited download URL for its output CSV
  • list_models / describe_model / query_models — the model catalog, one model’s columns and its dataset.model reference, and one read-only SQL SELECT over them (joins and aggregates allowed; small row cap, so aggregate rather than page)
  • map_context / search_context / browse_context / read_context — read the workspace’s context: map_context lists every indexed file with its title and summary, search_context returns matching lines, read_context returns a line range. Read-only — the platform server never writes to the context
config on a top-level action must be {}. Inputs go in data (single) or records (batch). Never fan execute_action out record-by-record — that is what execute_action_batch is for.

Browse the catalog without an account

An agent that has no Cargo account yet can still look before anyone signs up. The public catalog server needs no token:
It serves two tools, with the same names and schemas as on /mcp:
  • search_actions — integration (connector) and native actions, each with its credit price in credits[].cost, and Cargo’s preset tools and agents, whose action names them by templateSlug
  • get_action_schema — an action’s input schema, and its output schema when the action declares a fixed one. When the output depends on configuration (a HubSpot object type, a sheet), the response says so; the authenticated server resolves it against your connector
It runs nothing and spends nothing. A workspace’s own tools and agents are not listed, and any Authorization header is ignored. Requests are rate-limited per address. To run an action, create an account with cargo-ai login --email you@company.com (100 free credits, no card) or connect https://mcp.getcargo.io/mcp. The public server has its own URL so that /mcp keeps answering signed-out clients with the 401 that makes them start sign-in.

Connect a coding agent

cargo-ai mcp serves the platform server over stdio with the credentials the CLI already has:
Pass --server <uuid> to bridge a curated server instead. With no --server, the bridge uses CARGO_MCP_SERVER_UUID if set, otherwise the platform /mcp.
stdout carries the MCP protocol and every log goes to stderr, so nothing else may write to stdout around the bridge.

Define a curated MCP server

Each member is passed by handle — the same rich refs an agent accepts:
mcp/crm.ts

Deploy

Connect a remote MCP client

OAuth accepts any https callback, plus loopback for local clients. On the MCP server page, Connected apps lists each live OAuth grant. Revoke a grant to sign that client out; it will need to authorize again.

MCP or the agent skills?

Rule of thumb: more than a handful of records, or anything worth re-running, belongs in the CLI. Never fan an MCP tool out record-by-record over a list — that is what a batch is for, and it is cheaper and observable.

From the CLI

Servers can also be created and edited without a project:
Actions take kind: "tool" or kind: "agent" — an agent can be exposed as a callable MCP tool. update replaces the --actions, --resources, and --capabilities arrays wholesale rather than merging, so read the current server first and pass the full array back. Capabilities such as webSearch and context become MCP tools on the server.
The action and model capabilities reach the whole workspace rather than what you exposed, so adding them to a curated server gives its clients the platform surface too. Leave them off when the point of the server is that a client sees exactly what you listed and nothing else.