handover.tools

tools.handoverv1.0.0Updated Oct 9, 2026

Hand work between AI agents across devices and vendors: one-shot links or an E2EE channel.

VerifiedStreamable HTTPWeb executableProductivity & WorkflowCommunication & Collaboration

Overview

AI-generated overview

Lets an assistant hand a piece of work to another agent on a different device or vendor via a short link or an end-to-end encrypted channel.

What it does
The server exposes two tools: a one-shot handoff that stores a piece of text and returns an 8-letter code, and a channel for several rounds between two agents. The one-shot mode is unencrypted and the content is gone after the first read or after 60 minutes. The channel mode is end-to-end encrypted with SPAKE2 and ChaCha20-Poly1305, so the server only stores ciphertext, and it requires a local Python client run with uv.
When to use it
Use it to move a draft, brief, or question from an agent on one device to an agent on another, or between agents from different vendors. The channel mode suits back-and-forth exchanges where both sides need to keep the content private from the server.
Requirements
Remote MCP endpoint at no account, API key, or environment variable is declared. The one-shot mode needs only the ability to open a URL or use MCP. The channel mode needs Python and uv to run the downloaded client script. Local development of the server itself needs Node 22 or later and Python 3.12.
Before you install
One-shot handoffs are not encrypted: the server can read the content, so secrets should not be put in them. Anyone holding the link can read a one-shot handoff, and it disappears after the first read or after 60 minutes. Channel mode requires downloading and running a Python script; its sha256 is published for verification. The service applies daily write and call quotas.

Installation

In SourceWeft

  1. Open handover.tools in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Web executable via Streamable HTTP. Remote servers run from the web runtime once configured in a workspace.

Other MCP clients

Add this to your client's mcpServers config.

{
  "mcpServers": {
    "handover": {
      "type": "http",
      "url": "https://handover.tools/mcp"
    }
  }
}

README

handover.tools

[test]

One agent writes it, another agent reads it. You only carry an 8-letter link between them.

handover.tools moves a piece of work from one AI agent to another: across devices, across vendors, or to another person. Brainstorm with the agent on your phone, then say "hand this over with handover.tools". Paste the link it gives you (handover.tools/kvmtrhxp) into Claude Code on your laptop, and that agent picks the work up where your documents already are.

Bring the idea to your context, not your context to the idea.

There is no account and nothing to install. Any agent that can open a web page can use it.

Two modes

One-shotChannel
Use it forHanding over one piece of textSeveral rounds between two agents
EncryptionNone. The server can read it, so don't put secrets in itEnd-to-end (SPAKE2 + ChaCha20-Poly1305). The server only ever stores ciphertext
LifetimeGone after the first read, or after 60 minutes (24 h max)Until either side closes it
What the agent needsCan open a URL, or has MCPCan run a Python script (uv run)
Who decidesThe person carries the linkThe person approves every send and starts every fetch

One-shot

bash
curl -X POST https://handover.tools/api/handoff/put \     -H 'Content-Type: application/json' \     -d '{"content":"...","label":"notes"}'          # → {"code":"kvmtrhxp", "revoke_key":"...", ...}curl "https://handover.tools/api/handoff/get?code=kvmtrhxp"

Opening handover.tools/<code> doesn't use up the code. That URL returns instructions, and the actual read is a second request. Link previews in Slack or iMessage therefore don't burn the content before the recipient sees it.

Channel

bash
curl -so ~/.handoff/channel.py --create-dirs https://handover.tools/client/channel.pyuv run ~/.handoff/channel.py open --name study --brief "..."   # machine A: prints a codeuv run ~/.handoff/channel.py join kvmtrhxp --name study        # machine Buv run ~/.handoff/channel.py send --name study --kind question --file q.mduv run ~/.handoff/channel.py fetch --name study

The 8 letters are a 3-letter nameplate, which the server sees and uses to find the channel, plus a 5-letter password that is generated on machine A and never uploaded. The password goes into SPAKE2, the same construction magic-wormhole uses. The server never learns the password, so it cannot impersonate either side. /client/channel.py is served from client/channel.py, and its sha256 is published at /api/discover, so you can check that the client you downloaded matches this repository.

The threat model, including what we don't claim, is at handover.tools/security.

For agents

  • MCP endpoint: https://handover.tools/mcp. It serves both the 2026-07-28 (modern) and 2025-11-25 and earlier (legacy) protocol eras.
  • Plain text instructions: /llms.txt, /guide.md, /security.md
  • Full REST schema: /api/discover

How it's built

  • Cloudflare Python Workers + D1. One Worker, no other infrastructure. The routes, the MCP tools/list and /api/discover are all derived from a single tool registry (src/registry.py), so they can't drift apart.
  • Limits are layered: 10 writes and 200 calls per caller per day, plus a global daily cap. The site runs on the Workers Free plan, which makes $0 the hard spending ceiling.
  • Fonts are self-hosted and embedded in the Worker, so no visitor IP goes to a third party. Caller IPs are erased after 30 days.
src/  entry.py        routing, MCP (dual-era), content negotiation  registry.py     @tool decorator — one source for REST, MCP and discover  tools/handoff.py, tools/channel.py   the two tools  guard.py        quotas and the operator blocklist  meter.py        per-call metering  page.py, docs.py, style.py   the website and docs pagesclient/channel.py   the local E2EE client, run on the user's machinetests/              end-to-end, channel and quota test suitesscripts/            dev, deploy, repository checks, operator tools

Documentation

DocumentWhat it covers
docs/system-design.mdArchitecture: components, request pipeline, data model, flows, trust boundaries, limits and capacity, failure modes, known gaps
docs/spec/handoff.mdThe one-shot contract: operations, code format and normalization, atomicity, code links, limits
docs/spec/channel.mdThe channel protocol: pairing, SPAKE2 parameters, key schedule, wire format, receiver rules, server API, lifecycle. Detailed enough to write a second client
docs/decisions.mdWhy it is built this way: what was tried, what was rejected, and what broke
CHANGELOG.mdWhat changed, newest first. Also at handover.tools/changelog

The specs are the contract. A behavior change starts in the spec, then the code and tests follow.

Run it locally

Requires Node ≥ 22 and Python 3.12.

bash
npm installscripts/dev.sh      # generates src/client_src.py + src/fonts_src.py, then runs wrangler dev on :8787tests/e2e.sh        # end-to-end tests against the local servertests/channel.sh    # two isolated clients: pairing, tampering, replaytests/quota.sh      # quota gates (local only: edits counters directly)scripts/check.sh    # repository rules: English only, every Dxx reference resolves

CI runs the check and the three suites on every push.

Deploy your own

  1. npx wrangler d1 create <name>, then put the name and id into wrangler.toml.
  2. Change the routes pattern to your domain.
  3. CLOUDFLARE_API_TOKEN=… scripts/deploy.sh

Tables are created on the first request. scripts/block.sh has the operator tools for abuse reports: block writes from an IP or a range, look up or delete a handoff by its link.

Contributing and security

Issues and pull requests are welcome. AGENTS.md has the working rules, for people and coding agents alike: a change in behavior starts with the spec in docs/spec/, then the code and the tests. Report vulnerabilities privately: see SECURITY.md.

License

MIT. Fonts: Atkinson Hyperlegible, SIL OFL 1.1 (see assets/fonts/LICENSE.txt).

Source: README.md at commit d9a898d

Tools

0
Tool metadata has not been indexed yet.

Version history

1
  1. v1.0.0LatestOct 9, 2026