Drive Cargo from its hosted MCP server at https://mcp.getcargo.io/mcp — connect a client, discover and price an action, run it over one record or a batch, poll it, and read workspace models, with no CLI install. Also when to call an MCP tool instead of shelling out to `cargo-ai`. Triggers: "connect

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 ](../cargo ai/SKILL.md). 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 Warehouse SQL, aggregates, joins CLI ([ cargo storage ](../cargo storage/SKILL.md)) Build or edit a multi step workflow, tool, or play CLI ([ cargo orchestration ](../cargo orchestration/SKILL.md)) Workspace as code, plan and deploy CLI ([ cargo cdk ](../cargo cdk/SKILL.md)) Provision mailboxes, warm up, send CLI ([ cargo mailbox management ](../cargo mailbox management/SKILL.md)) 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. Thirteen tools cover discovering an action, running it, watching it finish, and reading data back. 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: 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. Any client taking a JSON block (Claude Desktop, Cursor, a project .mcp.json ): 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: The tool list is not fixed. The endpoint serves the platform tools below plus whatever that workspace published with defineMcpServer , so two tokens can see two different lists. Read the list you actually got rather than the one documented here. The spine For data: list models → describe model → query models . 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 carries a download for its output CSV. The tool description says never to loop it: take that literally. Spending before quoting. Run 10–20 records first, 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). query models mistaken for SQL. It lists records off one model with a limit and an offset. It does not aggregate, join, or filter by expression. Any question shaped like "how many", "grouped by", or "joined to" is a CLI question ([ cargo storage ](../cargo storage/SKILL.md)). Say so, rather than pulling rows and counting them yourself, which silently truncates at the limit. 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.

cargo-mcp