Switchboard

The agent that does the after-call paperwork the receptionist used to do.

A customer calls a service business and describes a job. The instant the call ends, Switchboard runs a gated multi-step DAG — Claude reads the messy spoken transcript, extracts a structured record, dedupes against existing customers, and (behind a hard gate on the irreversible write) files a clean, dated, queryable job record into Notion with a follow-up scheduled so the lead never goes cold.

The trigger is a live phone call — a real-world event, not text already inside a computer. The output is structured data the business keeps and queries — a real Notion row, not a chat transcript.

Built for the AI Agents for Productivity Hackathon · Notion + Anthropic tracks.

The reusable asset

This repo ships an open Notion Intake MCP server (switchboard/notion_mcp.py) exposing four tools over the Model Context Protocol:

Clone the repo, drop in your Notion token + database id, and point it at your own intake (call, email, or form) — working by Monday.

Related MCP server: Notion MCP Server

Quick start

# 1. Install
pip install -e .

# 2. Run the whole loop OFFLINE on a sample call transcript — no keys needed.
#    Claude extraction is stubbed deterministically and Notion is mirrored in
#    local SQLite, so a judge can verify the gated DAG end to end immediately.
python -m switchboard.agent_loop --demo

# 3. Run it GATED — the irreversible Notion write BLOCKS for approval, proving
#    the gate is real, not decorative.
python -m switchboard.agent_loop --demo --autonomy gated
# ...then approve the blocked run:
python -m switchboard.agent_loop --approve <run_id>

# 4. Run the test suite (in-memory, no keys).
python -m switchboard.tests

Going live

cp .env.example .env          # fill in NOTION_TOKEN, NOTION_DB_ID, ANTHROPIC_API_KEY
python -m switchboard.agent_loop --transcript path/to/call.txt --caller "+17875551234"

With keys present, extraction runs on Claude and the record is written to your real Notion database. Without keys, the same code path runs offline against the SQLite mirror — identical record shape, identical DAG.

Run the MCP server

python -m switchboard.notion_mcp           # serves the 4 tools over MCP (stdio)
python -m switchboard.notion_mcp --list    # print the tool catalog as JSON

The stdio server is built by notion_mcp.build_mcp_server() — the single source of truth for the served surface. python -m switchboard.tests includes test_mcp_server_advertises_tools, which builds that exact server and verifies, over the MCP protocol (not the static --list constant), that it advertises all four tools and that a real call_tool round-trips to the backend. So the stdio serving path is proven, not assumed.

Connect it to Claude Desktop (or any MCP client) — add to claude_desktop_config.json:

{
  "mcpServers": {
    "switchboard-notion": {
      "command": "python",
      "args": ["-m", "switchboard.notion_mcp"],
      "env": { "NOTION_TOKEN": "secret_...", "NOTION_DB_ID": "..." }
    }
  }
}

Omit env to run the server against the offline SQLite mirror with no keys — the four tools still appear in the client's tool list and work end to end.

The gated DAG

intake → extract → dedupe → [GATE] file → schedule → verify
 safe      safe      safe   irreversible   safe       safe

• SAFE stages run automatically on tick.

• IRREVERSIBLE stages (file — it writes to the owner's live workspace) BLOCK for approval when the run's autonomy is gated (the default for anything touching a customer's data). On --autonomy auto they run immediately.

• approve runs the blocked stage live, then continues the run forward through the remaining safe stages — the same gate shape Switchboard's orchestration engine uses in production.

Every stage transition is logged to a local SQLite runs blackboard, so the DAG is fully visible and queryable (and lights up on screen in the demo).

Watch it run

viewer/index.html is a zero-dependency page that polls the viewer/run.json snapshot the agent writes on every tick. Open it next to a live run and the DAG lights up stage by stage, the filed record fills in, and the header shows an instrumented time delta — elapsed_seconds vs. the hand-timed baseline_seconds (411s = 6:51), read straight from run.json by renderClock. The on-screen "6:51 by hand → seconds, 0 typing" number is the agent's own counter, not a desk stopwatch.

python -m switchboard.agent_loop --demo      # writes viewer/run.json
# then open viewer/index.html in a browser

Already captured: demo/ holds a screen recording (switchboard-demo.mp4) of this viewer animating through a real autonomous run and a real gated run (it blocks at the irreversible Notion write, then approves), plus the three frame screenshots and the exact run.json snapshots behind them. See demo/README.md — every asset is reproducible from the commands above.

Architecture

LIVE CALL ─► voice secretary layer ─► call-ended payload ─► Switchboard DAG ─► Notion Intake MCP ─► NOTION JOBS DB
                (reused)                {caller, transcript}    (gated)            (4 tools)         (the kept record)
                                                              Claude extracts +
                                                              plans tool calls

• Reasoning core: Claude reads the transcript, emits strict JSON, and plans the tool calls. The orchestration is a real gated DAG, not one mega-prompt.

• Persistence + payoff: the Notion Intake MCP server writes structured data into a real Notion database.

Note: Switchboard reuses an existing voice secretary intake layer and gated orchestration engine. This repo is the hackathon's net-new surface — the Notion MCP server, the agent loop, the lane definition, and the offline demo harness.

License

MIT — Hackathon submission. Clone it, aim it at your own intake.