Numbers
Plan quotas still apply.
create_agent succeeds when the workspace has no free agent slot, but the agent comes back with active: false and does not run until a slot frees up. See Plans and billing.
Scope
A key reaches one workspace, entirely. Every tool resolves its target through the same ownership check the app uses, so anagentId or a spaceId from another workspace comes back as Agent not found or Box not found, never as a partial answer.
There is no read-only key, no per-agent key and no per-tool permission. If you need a narrower blast radius, use a separate workspace.
What the API refuses to do
Accept a secret
Accept a secret
attach_connector declares a connector and stops there. API keys, OAuth sign-ins and passwords are entered by the account owner in the app, and they are stored on the Box, never in the Rerun database. plan_template_capture returns credential variable names only, never values.Publish a template
Publish a template
submit_template moves a draft to review. Publishing is the Rerun team’s decision. Payout onboarding needs a browser.Manage the account
Manage the account
No billing, no plan change, no subscription, no member or invitation management, no workspace creation, no Box creation. Create a Box in the app first.
Manage API keys
Manage API keys
You cannot mint or revoke a key over MCP. That would let a leaked key extend itself.
Reach agent memory or files
Reach agent memory or files
The API covers prompts, skills, schedules, triggers, connectors, databases and runs. Memory entries and workspace files are reached by the agent itself, or by you in the app.
Upload a file or generate an image
Upload a file or generate an image
add_template_image takes a public https URL and re-hosts the bytes. Produce the image elsewhere.Guarantees
- Destructive tools ask twice. Seven tools refuse to run without
confirm: true. See Destructive tools. - Skills are written atomically.
upsert_skillstages the whole folder and swaps it in, so a failed call never leaves a half-written skill. - Multi-statement SQL is transactional.
db_executeandspace_db_executewithstatementsapply all of them or none. - A tool error is readable. Failures come back as
isErrorresults with a message written to be acted on, not as an opaque protocol error. See Errors. - Renaming is safe. Changing an agent
namenever moves itshandle, so@handlereferences in other agents keep resolving.
Availability
The MCP endpoint and the webhook front door run on the Rerun app. The agents themselves run on your Box. A Box that is shut down or restarting makes agent-facing tools fail with a reachability error, while catalog tools likelist_connectors keep working.