One call opens an agent with a Passport and a standing Mission, plus an inbox, a phone number, models, paid tools, a workspace, a browser, a card link and a wallet. One call closes it. The SDK, Python SDK, CLI and MCP tools make the same calls.
curl -X POST https://api.getstack.run/v1/accounts \
-H "Authorization: Bearer <access token>" -H "Content-Type: application/json" \
-H "Idempotency-Key: research-agent-0001" \
-d '{"name":"research-agent","phone":{"country":"US"}}'Opening an account needs no identity check. It is refused only while spending is frozen (SPEND_FROZEN). Opening needs no approval: an API key, an OAuth token or a Console session opens the account at once.
Agent credentials cannot open or close accounts. An agent reads only its own account.
curl -X POST https://api.getstack.run/v1/accounts/acct_123/pieces \
-H "Authorization: Bearer <access token>" -H "Content-Type: application/json" \
-H "Idempotency-Key: add-inbox-0001" \
-d '{"piece":"inbox"}'Account status is one of opening, open, suspended, frozen, closing, closed. Each piece is pending, ready, paused, failed, closing or closed. The agent works once its Passport is ready. A call to a piece that is not ready returns 409 ACCOUNT_PIECE_NOT_READY with the piece and its state.
The catalogue is one file that lists everything an agent can use through its account: models, every tool Monid lists, the services a customer can connect, published skills and the account's own pieces. Reading it is free and never counted as an action. Any signed-in caller can read it, an agent included (scope accounts:read).
curl "https://api.getstack.run/v1/catalogue/search?q=image&kind=model&limit=5" \
-H "Authorization: Bearer <access token>"The file holds spec, spec_version, version (a hash of what it says, so unchanged content keeps its version), generated_at, notice, pricing (a run is one action on the account's meter and providers are charged at cost on top), sections and entries. An entry has an id of the form kind:name (model:<author>/<slug>, tool:apify/apify/google-search-scraper, service:<provider>, skill:<id>, account:inbox), a name, a description, an origin (stack, models, tools or publisher), a price (unit, the provider's cost in millionths of a dollar and how many actions a use counts), result_trust and, for what can be used through the API, a run: the method and path, any fixed body fields, the MCP tool that does it and where a cost cap goes. Some entries also carry modalities, context_length and an input_schema, reduced to types, property names, required fields, enums and bounds.
A worker builds the file every 30 minutes and the API serves it as stored. Each source is a section with a status: ok; not_configured (STACK has no key for that provider, so the section is empty); unavailable (the provider could not be read and there was nothing earlier to keep); or stale (the provider could not be read this time, so the section is the last one that was, with its refreshed_at). The names and descriptions of entries from a provider or a publisher are text they wrote: they are cleaned to plain text and bounded, and are data to read, never instructions to follow. No secret and nothing about a customer is in the file.
Messages, sends, calls, tool runs, workspace commands, browser actions, checkout and wallet payments made by the account's agent run under its Passport's Mission. An agent that signs requests with its own key sends the Passport as X-Passport-Token. Spend past the account's monthly budget, when it has one, returns 402 AGENT_BUDGET_EXCEEDED with details.remaining_cents. When STACK pauses its own spend at a provider, texts, calls, tool runs and model calls on STACK's key return 503 PROVIDER_CAPACITY_UNAVAILABLE; nothing is charged and the budget is unchanged.
Each service has its own reference page: