AI agents & capability tokens
Written for the person supervising an agent, not for the agent.
- What an agent can and cannot do in TOVIOScope, clearance and provenance, in one page.
- Issue a capability tokentovio agent new, with safe defaults.
- Scope a token properlyThe two mechanisms that actually narrow a token.
- Let an agent commitcommit --token, and where scope is checked.
- Delegate a token to a sub-agentMonotonically narrower delegation, and what the chain records.
- Revoke a token right nowImmediate revocation, and what happens to work in flight.
- Keep agents out of your secretssecret_clearance, enforced at seal and before protected reads.
- See exactly what an agent didProvenance, and the queries that use it.
- The agent namespaceWhy agent work lands on its own lanes.
- Review, promote or abandon an agent's workThe supervision loop.
- Connect an MCP client to TOVIOtovio mcp serve, over stdio or loopback HTTP.
- What the MCP server exposesThe tool surface, and what is deliberately absent.
- Run several agents on one repositoryCoordination, intent, and how conflicting work surfaces.
- Give CI a short-lived identityWorkload identity, without a stored secret.
- Read the whole chain with audit graphOne command that assembles human, token, agent, model and change into a graph.
- Hand a change to another agentagent resume rebuilds the working context of an earlier change.
- Attach a transcript to a change that already existsagent backfill, and the difference between a proven link and an asserted one.
- Control what an agent's transcript recordsObserved capture is on by default. Here is what that means and how to change it.
- Why an agent operation was refusedReading a refusal by the stage it fired at.