prog-strength-agent

User-facing agent service for the Prog Strength training tracker. Wraps Claude with the prog-strength-mcp tool layer so natural-language requests like "log Tuesday's push day" or "how has my bench progressed?" map to real reads and writes against prog-strength-api.

Architecture

user → frontend → POST /chat → AGENT ─→ Claude (tools=[…])
                                  │
                                  ├─→ MCP (list_workouts, list_exercises, create_workout)
                                  │       │
                                  │       └─→ API (Go) → SQLite
                                  │
                                  └─ injects authenticated user_id into every tool call

• Persistent MCP session. Opened at FastAPI startup via the lifespan handler, reused across all /chat requests. Reconnects on transport failure.
• Stateless /chat endpoint. The frontend sends the full message history each turn; the agent owns no per-conversation state.
• JWT-validated requests. The agent shares JWT_SIGNING_KEY with the API and trusts tokens minted by the API's OAuth callback. sub is the authoritative user_id — the agent strips user_id from the tool schemas it shows Claude and injects the authenticated value when forwarding to MCP, so Claude can never spoof a different user.

Configuration

Env var	Required	Default	Notes
ANTHROPIC_API_KEY	yes	—	Standard sk-ant-… key.
JWT_SIGNING_KEY	yes	—	MUST match the API's signing key. Used to verify inbound user JWTs.
PROG_STRENGTH_MCP_URL	no	http://mcp:8000/mcp	In prod this resolves over the shared docker network.
AGENT_HOST	no	0.0.0.0
AGENT_PORT	no	8001
CLAUDE_MODEL	no	claude-sonnet-4-6	Any Anthropic model id. Default is Sonnet 4.6 — strong tool-use behavior at a fraction of Opus pricing.
CLAUDE_MAX_TOKENS	no	2048	Per-turn cap.

Local development

uv sync

# Run the MCP server first (separate shell) — see prog-strength-mcp.
# Then point the agent at it. Both must share the same JWT_SIGNING_KEY.

export ANTHROPIC_API_KEY=sk-ant-…
export JWT_SIGNING_KEY=<dev key>
export PROG_STRENGTH_MCP_URL=http://localhost:8000/mcp

uv run prog-strength-agent

The service is then reachable at http://localhost:8001. Try:

curl http://localhost:8001/health

/chat

POST /chat accepts an Anthropic-format message list and returns the final assistant turn after the tool-use loop terminates:

curl -X POST http://localhost:8001/chat \
  -H "Authorization: Bearer $JWT" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": "what exercises do I have for chest?"}
    ]
  }'

The current implementation is non-streaming — the response arrives after the full tool-use loop completes. SSE streaming is a planned follow-up; it'll keep the same /chat shape but emit text_delta and tool_use events as they happen.

Deployment

Deployed on the same EC2 host as the API and MCP. The docker-compose service joins the shared prog-strength external network so it can resolve mcp and api by service name. Caddy reverse-proxies agent.progstrength.fitness → agent:8001.