Agent Handoff Protocol

Durable agent sessions that outlive their host.

A protocol and reference implementation for serializing a running agent's state, transferring it to a provisioned sandbox on a different machine, and metering what it costs to keep running there — built as an MCP server over a Postgres-backed state machine.

Live: agent-handoff-protocol.vercel.app · /dashboard shows a real transfer, end to end, on a live Neon database.

---

What this is

Long-running agent loops outgrow the machine they started on. This repo is the boring infrastructure for handling that gracefully:

1. snapshot_state captures a session's system prompt, message history, tool state, and MCP config (credentials as vault references, never raw secrets) into a Postgres row, along with a checksum of the snapshot.
2. provision_runtime sizes a destination sandbox and opens a fixed compute budget. Requires a transfer-authorization token.
3. push_state uploads the snapshot to the destination, optionally verified against the checksum from step 1.
4. activate boots the destination from the snapshot — the one irreversible step in the whole protocol. Also requires a token.
5. report_usage lets the destination's own metering daemon report spend against its budget, flipping the transfer to insolvent once it's exhausted. An hourly cron job auto-terminates any transfer left insolvent past its grace period.
6. get_status reads back the full transfer, budget, and ordered event log — this is what the dashboard renders.

No tool in this surface resembles "does the agent want to transfer." That decision belongs to whoever calls provision_runtime / activate — a human, a script, a scheduler — and per the auth model below, only that orchestrator can ever produce a valid token for those two calls. The full reasoning behind the boundary is in docs/DESIGN.md §5.

The landing page carries a short piece of narrative flavor text alongside the real, live event data — clearly labeled as fiction, not telemetry. The mechanism is real; the story is a showcase layer on top of it.

Architecture

flowchart LR
    subgraph Orch["Orchestrator (human/script)"]
        O[issueTransferToken]
    end

    subgraph Source["Source runtime"]
        A[Agent loop]
    end

    subgraph MCP["Transfer MCP server (packages/mcp-server)"]
        T1[snapshot_state]
        T2["provision_runtime (token)"]
        T3[push_state]
        T4["activate (token)"]
        T5[report_usage]
        T6[get_status]
    end

    subgraph Core["@ahp/core"]
        SVC[service.ts state machine]
        DB[(Neon Postgres via Drizzle)]
    end

    subgraph Dest["Destination runtime"]
        D[Resumed agent loop]
        M[Metering daemon]
    end

    subgraph Web["@ahp/web on Vercel"]
        DASH[/dashboard/]
        CRON["/api/cron/reap (hourly)"]
    end

    O -.mints token, never via MCP.-> T2
    O -.mints token, never via MCP.-> T4
    A -->|calls| T1 & T2 & T3 & T4
    T1 & T2 & T3 & T4 & T5 & T6 --> SVC --> DB
    T4 -.boots.-> D
    M -->|calls| T5
    DASH -->|reads| DB
    CRON -->|reaps expired + insolvent| DB

Repo layout

packages/
  core/         Drizzle schema + framework-agnostic service layer (the state machine, auth, tests)
  mcp-server/   MCP stdio server exposing the six tools above, wraps @ahp/core
  web/          Next.js app: landing page + /dashboard (live) + /docs + /disclaimers + cron route
scripts/
  demo.ts       Runs one full lifecycle end-to-end against a real Neon DB
docs/
  DESIGN.md     Full technical spec, including what's simplified for this showcase
  ROADMAP.md    Phased plan for what's built vs. what's next, review-approved
  TEAM.md       Named draft-only personas — see for the "no auto-publishing" hard rule
content/
  drafts/       Where personas draft content; nothing here is published automatically
.github/workflows/ci.yml   Build + typecheck + test on every push/PR to main

Three packages, one schema — the MCP server, the demo script, and the dashboard's read queries all call the same @ahp/core functions rather than reimplementing the state machine three times.

Quick start

git clone https://github.com/zordhalo/agent-handoff-protocol
cd agent-handoff-protocol
pnpm install
pnpm --filter @ahp/core build   # @ahp/core ships compiled (dist/ is gitignored); demo.ts and the web app both import it

# Pull DATABASE_URL, TRANSFER_TOKEN_SECRET, CRON_SECRET from the Vercel project
vercel link
vercel env pull .env.local

pnpm db:migrate      # apply the schema to your Neon DB
pnpm demo            # run one full staged→provisioned→pushed→active→insolvent→terminated cycle
pnpm --filter @ahp/web dev   # open http://localhost:3000/dashboard to see it

Auth setup

provision_runtime and activate both require a transfer-authorization token (docs/ROADMAP.md Phase 1 item 4). Tokens are HMAC-signed, short-lived, and single-use, and are minted by issueTransferToken from @ahp/core — never by an MCP tool, so the source loop (the agent) has no path to mint one itself. scripts/demo.ts plays the orchestrator role and mints its own tokens; a real deployment would do this from whatever process is actually driving the handoff (a script, a human-triggered API route).

# TRANSFER_TOKEN_SECRET gates provision_runtime/activate.
# CRON_SECRET gates the /api/cron/reap route (Vercel attaches it automatically
# to its own scheduled invocations once it's set as a project env var).
node -e "console.log(require('crypto').randomBytes(32).toString('base64url'))"

Set the output as TRANSFER_TOKEN_SECRET (and a separately generated value as CRON_SECRET) in your Vercel project's environment variables, then vercel env pull .env.local again to pick them up locally.

Running the MCP server against a real agent client

pnpm --filter @ahp/mcp-server build

Point your MCP-capable client at packages/mcp-server/dist/index.js (stdio transport) with DATABASE_URL set in its environment.

Database setup

This repo assumes Neon Postgres, provisioned through Vercel's integration marketplace (Project → Storage → Neon), which sets DATABASE_URL for you. Any Postgres connection string works — @ahp/core only needs it in the environment.

pnpm db:generate   # regenerate drizzle/ migrations after a schema change
pnpm db:migrate     # apply them

What's real vs. simplified

This is a working reference implementation, not a hardened production system — the "destination runtime" in the demo is a script writing to the same database the dashboard reads, not an isolated sandbox, and credRef is still a free-text string rather than a real vault lookup. Auth-gating and the insolvency-termination lifecycle, previously listed as gaps, are now real (docs/ROADMAP.md Phase 1). The current, honest breakdown of what's real vs. simulated is docs/DESIGN.md §7, and what's planned next is docs/ROADMAP.md.

Stack

• Neon Postgres, provisioned via the Vercel marketplace
• Drizzle ORM with the @neondatabase/serverless HTTP driver
• @modelcontextprotocol/sdk for the tool server
• Next.js App Router, deployed on Vercel, incl. a Vercel Cron route
• Vitest for integration tests against a live Neon database
• pnpm workspaces monorepo, GitHub Actions CI

License

MIT — see LICENSE.