ob1-local-http

# OB1 Local HTTP

## Problem

The canonical Open Brain stack assumes Claude Code can talk to a remote
Supabase MCP server. In environments that disable MCP entirely -- corporate
networks, air-gapped offices, restricted Claude Code builds -- that path is
not available. This skill replaces it with `curl` calls to a LAN-resident
Open Brain (see the `local-brain-no-mcp` recipe), keeping the same
capture-and-recall behavior without any MCP protocol involvement.

## When to Use

- The user wants to remember, record, save, capture, or note something.
- The user wants to search, recall, retrieve, find, or look up something
  they previously captured.
- The user wants to see recent thoughts (e.g. "what have I been thinking
  about", "show me today's notes").

## When Not to Use

- The environment has a working remote Open Brain MCP connection -- prefer
  the canonical MCP-based capture/search tools.
- The required environment variables `BRAIN_URL` and `BRAIN_ANON_KEY` are
  not set on the dev host -- this skill cannot function without them; ask
  the user to follow `local-brain-no-mcp`'s install instructions first.

## Required Environment

On each dev host that will use this skill, the user must export:

```sh
export BRAIN_URL="http://<brain-host>:8000"   # Supabase Kong gateway
export BRAIN_ANON_KEY="eyJhbGciOi..."          # written by setup.sh
```

If either is missing, stop and tell the user. Do not guess values.

## Process

### Capture

When the user says something like "remember X" or "save this thought":

```sh
curl -fsS -X POST "$BRAIN_URL/functions/v1/capture" \
  -H "apikey: $BRAIN_ANON_KEY" \
  -H "Authorization: Bearer $BRAIN_ANON_KEY" \
  -H "Content-Type: application/json" \
  -d '{"content":"<the thought>","metadata":{"source":"claude-code"}}'
```

Optional `metadata` fields: `source`, `tags` (array of strings),
`thread_id`, anything else useful. The server fingerprints content and
de-duplicates -- re-capturing identical text returns the existing id.

### Search

When the user wants to recall:

```sh
curl -fsS -X POST "$BRAIN_URL/functions/v1/search" \
  -H "apikey: $BRAIN_ANON_KEY" \
  -H "Authorization: Bearer $BRAIN_ANON_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query":"<what to find>","match_count":10,"match_threshold":0.65}'
```

`match_threshold` defaults to 0.7. Lower it to 0.5-0.65 for broader recall;
raise to 0.8+ for precision. Cap `match_count` at 100.

### Browse recent

When the user asks "what have I been thinking about" or wants a list rather
than a similarity search:

```sh
curl -fsS "$BRAIN_URL/functions/v1/list?limit=20" \
  -H "apikey: $BRAIN_ANON_KEY" \
  -H "Authorization: Bearer $BRAIN_ANON_KEY"
```

## Output

- For captures: confirm the id and the de-dupe fingerprint to the user in
  one sentence. Don't paraphrase the captured content back at them.
- For searches: surface the top results with similarity scores and
  created_at timestamps. Order by similarity descending. If no results
  cross the threshold, say so plainly and suggest lowering it.
- For browse: a compact bullet list with truncated content (first ~120
  chars) and timestamps.

## Failure Modes

- HTTP 502 with "unable to reach Ollama": the brain host's Ollama service
  is down. Tell the user to `docker compose ps` on the brain host.
- HTTP 502 with "did you 'ollama pull...'": the embedding model isn't
  loaded. Tell the user to run the documented `ollama pull` step.
- HTTP 502 with "embedding-dim mismatch": the brain's `EMBED_DIM` env was
  changed after the volume was initialized. Tell the user to read the
  "one-way door" section of `local-brain-no-mcp`'s README.
- HTTP 401 or 403: `BRAIN_ANON_KEY` is wrong or expired. Tell the user to
  check `supabase-docker/docker/.env` on the brain host.
- Network timeout: the brain host is unreachable. Tell the user to ping
  the brain host from this dev host.

## Notes

- Never log or echo `BRAIN_ANON_KEY` -- it's a long-lived JWT.
- This skill never installs or invokes any MCP server, by design.
- The brain host generates embeddings itself; this dev host doesn't need
  Ollama or any model files.