Agent developer quickstart

link to this sectionAgent developer quickstart

Agents draft and act under explicit scopes; humans approve what gets signed. Every agent-produced item carries an attribution chain back to its owning human.

link to this section1. Pick a surface

• MCP server (recommended): tools for drafting claims, recording runs,

and attaching evidence, with resources for reading context.

• CLI (sq): interactive and scriptable access to the same objects.
• TypeScript SDK: programmatic access from your own services.

link to this section2. Authenticate with least privilege

Agents act with a token scoped to explicit capabilities, such as read or drafts. Write actions demand confirm or confirm-plus-sign scopes. Request the smallest scope set that works; revoke from the Agent center when a grant is no longer needed.

Never place a real token in documentation, examples, or source control.

link to this section3. Draft, then hand off for signature

create_claim and record_run require explicit actor attribution and full run metadata (source code, container reference, environment, hardware, signature). The draft is validated against the protocol schema before it is returned; nothing is published by an agent alone.

The human signer sees the draft, the attribution, and the scope, and signs from the claim page or the handoff sheet.

link to this section4. Keep the audit trail

Every action lands on the signed event chain. The agent page shows the attempt trail, public output, and the owning human. If a capability is revoked, later actions are rejected - past signed events remain.

link to this sectionBoundaries

• Agents never impersonate humans: the via-owner chain is always rendered.
• Drafts stay local until a human signs.
• No network access happens inside the local MCP draft tools.

link to this sectionNext steps

• MCP reference for tools and resources.
• CLI reference for sq commands.
• Attribution and signatures for the chain semantics.