Intent Layer

Team memory for AI coding agents. Captures the intent behind every AI-assisted code change so the next session — yours or a teammate's — knows what was decided, what was rejected, and why.

Install

npm install -g @anupamvijayvergia/intent-layer

That's it. There are no API keys to configure, no Supabase URL to set, no .env to fill in. Everything that needs a secret runs on the dashboard you log in to.

Get started

Three commands, in this order:

intent-layer login        # opens browser, authorizes this device
intent-layer install      # registers a user-scope OS service that
                          # starts the watcher at login
intent-layer status       # confirms the daemon is running

After intent-layer install, the watcher runs in the background and auto-starts whenever you log in to your machine. There is nothing else to do — every Claude Code or Codex session you run will have its intents extracted and made available to your team's dashboard and to the MCP server.

What the watcher does

The background daemon watches your Claude Code (~/.claude/projects) and Codex (~/.codex/sessions) JSONL session files. When a session finishes (either a 60-turn batch or 30 minutes of idle time), it:

1. Chunks the session client-side (no LLM cost — pure text transform).
2. POSTs each substantial chunk to your dashboard at /api/v1/sessions/extract with your device's bearer JWT.
3. The dashboard runs the extraction LLM, persists the resulting intent memory scoped to your tenant, and returns the parsed payload.
4. Every 5 minutes the watcher pings /api/v1/sessions/touch so the Devices page on the dashboard knows this device is alive.

The customer-side runtime never holds a Supabase service key, an LLM API key, or any other infrastructure secret. After intent-layer login your ~/.intent-layer/config.json contains only:

{
  "jwt": "...",
  "user_id": "[email protected]",
  "org_slug": "your-team",
  "device_name": "your-laptop",
  "server_url": "https://your-dashboard.example.com"
}

Day-to-day commands

| Command | What it does | | --- | --- | | intent-layer status | Prints config and daemon health (running / stopped / failed). Pass --json for machine-readable output. | | intent-layer logs | Tails the daemon's platform-specific log (LaunchAgent log on macOS, journalctl on Linux, Windows Event Log). | | intent-layer list | Lists recent intents for the last 30 days via the dashboard's read API. | | intent-layer health | Confidence distribution + evolution counters. Read-only summary. | | intent-layer uninstall | Reverses install — removes the OS service. Your config and stored intents stay. | | intent-layer reset-local | Clears the watcher's local sync-state file so a re-run will re-process recent sessions. Does NOT touch the database. |

Per-repo setup (optional)

To enable the MCP get_file_intent / search_intent tools inside a specific repo (so Claude Code/Cursor/Codex can query intent before editing files), run from the repo root:

intent-layer setup

This writes .claude/CLAUDE.md, a PreToolUse hook, and .mcp.json into the repo so any AI agent that opens the repo automatically gets the intent-aware behavior.

Troubleshooting

• Error: not logged in. Run intent-layer login first. — your device JWT expired or was revoked. Re-run intent-layer login.
• Watcher not picking up sessions — check intent-layer status for daemon state and intent-layer logs for errors. Confirm the files actually live under ~/.claude/projects or ~/.codex/sessions.
• Verbose logging — set INTENT_LAYER_DEBUG=true in the daemon's env (launchctl setenv on macOS, systemd drop-in on Linux). This is the only env var the watcher reads at runtime; everything else flows from the dashboard.
• Switch dashboards — re-run intent-layer login and pick the new server when prompted; the JWT and server_url get rewritten atomically.

What's NOT here

The watcher is a thin client. It does not:

• Hold any database or LLM credentials. Compromising a dev laptop doesn't leak service-role access.
• Run background LLM calls on your machine (every extraction goes to the dashboard).
• Store intents locally in a way the rest of your team can't see.

If you're operating the dashboard side, see intent-layer-dashboard/README.md (separate package) for deployment and configuration.