Work with an agent

An agent can do more than answer questions about your notes. It can follow related records, compare choices, draft a plan and update the files that keep that work useful after the conversation ends.

Connect the agent to this brain

The brain template prepares two complementary pieces:

• Tools let a compatible agent search, read, add and update records through the MCP server.
• Skills guide workflows such as onboarding and reviewing captured notes.

Your client must load the tools and discover the skills. The template includes project-scoped MCP configuration, but clients differ in how they read and approve it. Registration alone does not prove that the current conversation is using the intended brain. Ask the agent to read a known record and check that it matches the file in your repository.

Use the MCP registration reference on GitHub (external site: github.com) for client settings, and the skill discovery reference (external site: github.com) when slash workflows do not appear.

Ask for an outcome and its evidence

For a fictional Odyssey brain, a useful request might be:

Review the departure preparations. Show which records support the plan and which questions still need a decision.

This gives the agent work to do and a way for you to assess it. Follow the source links, inspect the proposed changes and correct the assumptions before asking it to act on the plan.

Use guided workflows for repeated work

/brain-init helps propose a structure for your own brain. /process-notes helps review the capture inbox. These workflows depend on the agent client and its skill setup; MCP tools and slash commands are separate capabilities.

You can also use ordinary conversation to ask for a comparison, a draft or a specific file update. Keep requests bounded enough to review the result.

Know where judgment and permissions sit

Agents use their own model accounts or credentials. Their answers need review, especially when a source is old, incomplete or contradictory. The agent can operate on files, so its permissions matter too. Ordinary supported edits may be allowed directly; gated operations can ask for approval. Read what an approval will authorize before accepting it.

The rich interface makes decisions, sources and work steps visible alongside the conversation.

Let an agent use Apify, within limits you set

Apify runs paid web-extraction Actors. It is off until your governance policy (context/policies/governance.md) sets apify.enabled: true and the server has an APIFY_TOKEN in its own environment. The token never appears in the interface and is never typed into it. Settings › Accounts › External services › Apify shows the charged account and everything below in one place.

• Catalog search lets an agent look up Actors. Only the search words or the chosen Actor’s identifier are sent; finding an Actor never lets it run.
• Running Actors is separate. Ask shows you each exact run: the Actor, its build, what is sent and where, the account charged, the cap and anything not covered. You allow it once or deny it. Allow skips that review only for an approved Actor at its exact build, input schema and price.
• Approved Actors are grants you add in the policy file. Revoke or remove one in Settings; an approved run never adds one.
• Limits in apify.limits bound each run, turn, conversation and day. Without them nothing can run. Your overall budgets still apply, and the lower limit wins.

Every change you make in Settings opens the same policy review as an agent’s proposal, and nothing changes until you approve that exact diff. Turning Apify off stops new work but keeps grants and unresolved runs visible.

Spend appears in Settings, in Activity and in each turn’s cost details, apart from inference. A run shows as “up to” its reservation until Apify reports its charges; a start Apify never confirmed, or spend the host could not read, shows as unknown, never as $0. Stopping a turn sends an abort, but the run counts until Apify confirms it. Results stay readable here for the policy’s retention period; Apify may bill storage separately. Activity in Apify Console or other applications on the same account is not counted. The supported inputs (external site: github.com) list every field, and the CLI reference (external site: github.com) covers brain apify.