Relay

Live demo: https://relay-agent.fly.dev — the root redirects to a dashboard of real agent runs (cost, latency, outcomes).

An AI support-triage agent, built as a production service — not a notebook.

Relay receives support tickets for a fictional SaaS product (Lanekeep) over a REST API and works each one autonomously: it looks the customer up in a real database, classifies the ticket, searches the product documentation so every claim is grounded, and either sends a resolved reply or escalates to a human with a structured handover — streaming its reasoning steps to the client as server-sent events the whole way.

The agent loop is written by hand on the Claude API (no orchestration framework), so the control flow, step caps, and event stream are fully visible and testable.

Status / roadmap

• Phase 1 — Core agent service: FastAPI + SSE, hand-written agent loop, tools (lookup_customer, search_docs, set_category, send_reply, create_escalation), SQLite with seed data, keyword doc search

• Phase 2 — Guardrails: Pydantic-validated tool inputs, per-run cost budget with hard abort, write-tool policy (?dry_run=true), structured error events on API failure, per-step usage events with running cost

• Phase 3 — Evaluation harness: 12-ticket golden dataset, deterministic action/category grading plus LLM-as-judge grounding checks, JSON report artifact, threshold exit code for CI (python -m relay.evals)

• Phase 4 — Observability: JSON structured logs, OpenTelemetry spans per run/model-call/tool (OTLP export via OTEL_EXPORTER_OTLP_ENDPOINT), per-run metrics in SQLite, /metrics aggregates, /dashboard page

• Phase 5 — MCP server: the same tool registry (plus ticket lifecycle tools) served over the Model Context Protocol via stdio, behind the same validation and write-policy guardrails

• Phase 6 — Ship it: Dockerfile with container healthcheck, GitHub Actions CI (lint + tests + Docker smoke test on every push; on-demand eval workflow with report artifact), Fly.io deploy config

See docs/PROJECT_BRIEF.md for the full project definition.

Related MCP server: Customer Support MCP Server

Quick start

python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

cp .env.example .env   # add your ANTHROPIC_API_KEY

uvicorn relay.main:app --reload

Then, in another terminal:

./scripts/demo.sh

You'll see the agent's run streamed as SSE — text updates, each tool call and its result, and a final resolution event.

API

MCP server

The same tools are exposed over the Model Context Protocol, so Claude Desktop, Claude Code, or any MCP client can drive Relay directly:

claude mcp add relay -- /path/to/.venv/bin/python -m relay.mcp_server

Tool calls go through the same guardrail chain as the agent loop (Pydantic input validation + write policy). Set RELAY_MCP_ALLOW_WRITES=false to serve a read-only surface.

Tests

pytest

Tests cover the tools and the HTTP surface without calling the Claude API, so they run free and fast in CI.

Architecture

client ──POST /tickets/{id}/process──▶ FastAPI ──▶ agent loop (Claude API)
   ◀───────── SSE: text / tool_use / tool_result / resolution ─────────┘
                                          │
                          tools ──▶ SQLite (customers, tickets,
                                    escalations, replies)
                                └─▶ kb/*.md (docs search)

Deployment

CI runs lint, the 37-test suite, and a Docker build + container smoke test on every push. The eval suite runs on demand (Actions → Evals) against the ANTHROPIC_API_KEY repository secret and uploads the JSON report as an artifact.

Deploy to Fly.io with the included fly.toml:

fly launch --no-deploy
fly volumes create relay_data --size 1
fly secrets set ANTHROPIC_API_KEY=sk-ant-...
fly deploy

Or run the container anywhere:

docker build -t relay .
docker run -p 8000:8000 -e ANTHROPIC_API_KEY=sk-ant-... relay

License

MIT