Cargo — the hosted MCP server

Cargo has two surfaces. The rest of this bundle documents the CLI. This one documents https://mcp.getcargo.io/mcp, and, more usefully, when to reach for which.

Three different things here are called MCP. This skill is the hosted server Cargo runs, which you point a client at. Publishing a curated server out of your own workspace (ai mcp-server create, then cargo-ai mcp over stdio) and attaching somebody else's server to a Cargo agent (release update-draft --mcp-clients) are both cargo-ai. Check which one the user means before answering: the words are identical and the answers share nothing.

Which surface

The job	Surface
Run one action, or one action over many records	either; MCP if it is already connected
Find what Cargo can do, and what it costs	either (search_actions is the MCP half)
Read records off a model	either
One SELECT over models (aggregates, joins), 50 rows max	either (query_models is the MCP half)
Full SQL result sets, exports	CLI (cargo-storage)
Build or edit a multi-step workflow, tool, or play	CLI (cargo-orchestration)
GTM as Code: project plan and deploy	CLI (cargo-project)
Provision mailboxes, warm up, send	CLI (cargo-mailbox-management)
Segments, connectors, content libraries, alerts, hosting, billing admin	CLI
No shell at all (ChatGPT, Claude Desktop, claude.ai, n8n)	MCP, and say plainly what is out of reach

The rule underneath the table: MCP is the runtime, the CLI is the platform. Seventeen tools cover discovering an action, running it, watching it finish, reading data back, and reading the workspace context. Everything that builds something reusable is CLI only. An agent holding both should prefer the CLI for anything the user will want to re-run or version, and MCP for one-shot execution inside a conversation.

When the job routes to the CLI, this is the whole bootstrap:

npm install -g @cargo-ai/cli
cargo-ai login --email you@company.com   # emailed code, no browser; creates the account on first use
cargo-ai whoami                          # confirm the workspace before anything that spends

Connect

The endpoint is https://mcp.getcargo.io/mcp, Streamable HTTP. An unauthenticated request returns 401 with a WWW-Authenticate challenge carrying resource_metadata, so an OAuth-capable client discovers the authorization server, registers, and prompts the user with no configuration beyond the URL. A 401 on first connect is the handshake, not a fault.

claude mcp add --transport http cargo https://mcp.getcargo.io/mcp

Any client taking a JSON block (Claude Desktop, Cursor, a project .mcp.json):

{
  "mcpServers": {
    "cargo": {
      "type": "http",
      "url": "https://mcp.getcargo.io/mcp"
    }
  }
}

For CI, a headless agent, or a client with no OAuth, pass a workspace-scoped API token from Settings > API instead. Read it from the environment; never inline the value:

{
  "mcpServers": {
    "cargo": {
      "type": "http",
      "url": "https://mcp.getcargo.io/mcp",
      "headers": { "Authorization": "Bearer ${CARGO_API_TOKEN}" }
    }
  }
}

This endpoint serves the platform tools only. A curated server a workspace published with defineMcpServer is a different endpoint, at its own /v1/ai/mcpServers/<uuid>/mcp URL. Every token sees the same platform tools here; the token's permissions decide which calls succeed and which actions search_actions returns.

The spine

whoami                 → which workspace am I in, how many credits
search_actions         → find the action, and read its cost
get_action_schema      → what inputs it takes
autocomplete_action    → resolve a field needing a picked id (HubSpot object type, Slack channel)
execute_action         │ one record
execute_action_batch   │ many records
get_run / get_batch    → poll while outcome is "executing"

For data: list_models → describe_model → query_models. For the workspace context (read only): map_context, search_context, browse_context, read_context. Alongside, list_runs lists recent ad-hoc runs, and get_usage breaks the last 7 days of credit spend down by integration.

Open every session with whoami. The token binds the session to exactly one workspace and there is no flag to override it. A session pointed at the wrong workspace returns plausible, confidently wrong reads: the models are real and the records are real, they just belong to somebody else. Name the workspace back to the user before acting on anything.

search_actions prices the work before you do it. Each result carries credits[].cost beside the action object you pass verbatim to everything downstream:

{

**Pass that `action` object exactly as it comes, with no `config` key.** Inputs
belong in `data` (single) or `records` (batch) — never in a `config`, which is a
*node's* configuration and has no meaning on a top-level action. Inputs
misplaced there are silently dropped and the action runs with none, so an
unexplained empty result is worth checking against this first.

`get_action_schema` takes the same pair: the action, plus an optional `data` for
the actions whose output depends on their inputs — a HubSpot object type or a
target sheet decides which fields come back. The CLI's
`orchestration action get-output-schema` behaves identically.

Four `kind` values come back: `connector` (a third-party integration), `native`
(a built-in platform operation), `tool` (a saved workflow in this workspace),
and `agent` (an AI agent in this workspace). Narrow a noisy catalog with the
`kind` and `integrationSlug` filters.

## Three ways this goes wrong

**Fanning out `execute_action`.** One call per record is slower, bills more, and
leaves nothing to inspect afterwards. `execute_action_batch` takes the same
`action` plus a `records` array, produces one batch object, and a finished batch
returns one page of its output CSV in `data.text` (output columns prefixed
`output_`, 100 rows; while `truncated`
is true, call `get_batch` with the next `offset`). Read `data.text`, do not
fetch the URL. The tool description says never to loop
it: take that literally.

**Spending before quoting.** Run **10–20 records** first (`execute_action_batch`
takes a `limit` that submits only the first N), report the observed
cost and hit rate, then quote the full **record count** and **credit estimate**
and let the user approve. Hit rates on people data run 40 to 70 percent, so cost
per *usable* row is not the sticker price and is not knowable without the
sample. Full discipline:
[`../cargo-gtm/references/cost-discipline.md`](../cargo-gtm/references/cost-discipline.md).

**Paging rows out of `query_models`.** It runs one read-only SQL SELECT over
`dataset.model` references (take the exact reference from `describe_model`),
joins and aggregates allowed, and returns at most 50 rows (default 20). Answer
"how many", "grouped by", or "joined to" with the aggregate in SQL, never by
pulling rows and counting them yourself, which silently truncates at the cap. A
full result set is a CLI job ([`cargo-storage`](../cargo-storage/SKILL.md)).

## Anything that touches a person

The consent rules do not relax because the surface changed. A lawful basis, a
suppression check, and relevance to that person's job gate every step that
sources, enriches, or contacts someone. Bulk unsolicited messaging, purchased or
scraped lists, and consumer targeting are refused. The full text is
[`../cargo-gtm/references/acceptable-use.md`](../cargo-gtm/references/acceptable-use.md);
where no sibling skill is installed, the paragraph above binds on its own.

## Reporting back

Narrate and summarize; never paste raw JSON at a user. After a batch, give the
record count, the hit rate, the credits actually spent, and the download, in
that order.