Tools and permissions
Choose which tools your agent can call, pin the revisions you reviewed, and limit attempts per session.
Choose what your agent can use
- Tools: make only selected tools callable through Smoothie.
- Revisions: require the binding you reviewed; changed registrations need fresh review.
- Attempts: set a per-session call limit, not a global spending budget.
Your agent cannot add tools or raise that limit through call arguments. Grants allow repeated valid calls; they are not human approval of each call's arguments.
1. Review a local tool
Inspect the program and its input contract before granting access:
smoothie tool list --registry /absolute/tools.json --project /absolute/my-product --format json
smoothie tool show <id> --registry /absolute/tools.json --project /absolute/my-product --format jsonThese commands do not execute tools or read credential values. An SDK Tool is not automatically registered here. Terminal tool invoke plans first; execution requires --run --expect-revision.
2. Set access for the connection
Configure your agent to launch the reviewed tool with a limit of five attempts:
smoothie mcp serve --project /absolute/my-product \
--registry /absolute/tools.json --allow-tool <id>@<reviewed-revision> \
--max-tool-calls 5Replace the placeholders with the reviewed ID/revision; repeat --allow-tool for more tools. To change or remove local grants, end the connection, edit its launch arguments, then reconnect. Tools requiring per-call human approval remain blocked.
Both modes support model-free use and keep the same permission checks. Follow host setup to connect, Context and tool discovery for search, or Connect external tools for existing MCP servers.
Know the limits
Grants constrain Smoothie calls, not your agent's own shell/file access. Programs run with your OS permissions. This is not a sandbox; keep provider-side authorization and idempotency checks.
Do not automatically retry a missing, cancelled, or malformed reply: the tool may already have changed something. Check the outcome and cleanup failures first. Treat returned content as untrusted.