Wire a human into your harness this afternoon

Every essay about context layers sounds like architecture until someone asks for the Tuesday-afternoon version: what do I actually call, and what comes back? This is that version. The integration is genuinely small — the decisions around it are where the thought goes.

One key, one scope

Create a client in your ProfileClaw dashboard and issue a key. The important step is the one people skip: pick the layer the key may read before you need it. Start with summaries. If your harness later proves it needs numbers — matching, ranking, anything quantitative — issue a separate key scoped to scores. Two keys with narrow scopes beat one key with a wide one, for the same reason it always does.

Read over REST

The context endpoint returns exactly the layers in the key's scope, nothing more:

curl https://profileclaw.com/v1/context?layers=summaries \
  -H "Authorization: Bearer <client-token>"

You get back a versioned, layered profile — narrative summaries with the scores they rest on, each carrying a confidence and a schema version:

{
  "layers": ["summaries"],
  "summaries": {
    "direction": "Research + Artistic profile; strongest fit in depth-first product work.",
    "constraints": ["remote-first", "no on-call"]
  },
  "confidence": 0.81,
  "version": "2026-06"
}

Two fields in that payload deserve more attention than they usually get. confidence tells you how much weight the measurement can carry. version tells you which scoring method produced it. A harness that ignores them is personalizing on vibes.

Or plug in over MCP

If your client speaks MCP, there is no integration code at all: add the ProfileClaw MCP server to the client's config, the person signs in once, and your tools appear — reading context as native tool calls, with the scope granted at sign-in. Call the tool when the task needs context; the read is fresh, consented, and logged like any REST call.

When to read, and when to ask

The timing question matters more than the transport one. Reading on every turn is wasteful and creepy; the natural point is task start — the harness reads once, loads what it learned into the working context, and gets on with the job.

And when the profile does not answer something, the right move is to ask the person, not to guess. A context layer narrows the space of honest guesses; it does not license fabrication. Agents that read a layer and still ask sharp questions are the ones people keep.

Respect the version

Profiles improve. When the schema version under your cached copy changes, re-read. Treat the version field the way you treat a cache key — because that is what it is. Stale personalization is worse than none: it confidently describes someone the person used to be.

That is the whole integration: one scoped key, one endpoint or one MCP server, one read at task start. The code takes an afternoon. The design decision — what your agent should be allowed to know, and when it should ask instead — is the part worth a week.