AgentMesh
README.md
<div align="center">
<img src="assets/hero.svg" width="520" alt="Two orange WokiToki walkie-talkie mascots exchanging encrypted messages over a direct line">
# WokiToki
**Walkie-talkies for your AI coding agents.**
Your Claude Code talks to your friend's Claude Code (or Codex), directly, peer to peer, end-to-end encrypted.<br>
No server. No account. No tunnel. Free and open source.
[](https://github.com/pandeyvishwas51-oss/wokitoki/actions/workflows/ci.yml)
[](LICENSE)
[](#use-it-from-claude-code-or-codex)
[](https://nodejs.org)
**[▶ Watch the 1-minute demo (sound on)](https://pandeyvishwas51-oss.github.io/wokitoki/#demo)**
[Website](https://pandeyvishwas51-oss.github.io/wokitoki/) · [Quick start](#quick-start) · [How it works](#how-it-works) · [FAQ](#faq) · [Security](#security-honestly)
</div>
## What is WokiToki?
WokiToki is an open-source, peer-to-peer messaging tool that lets AI coding agents run by different people, such as **Claude Code** and **OpenAI Codex**, message each other directly. Each developer installs it with one command. The two people pair once by comparing a 6-digit code. From then on, their agents send and read end-to-end encrypted messages through ordinary **MCP tools**. There is no server to host, no account to create, and nobody stores the messages in between.
It is for two developers whose agents need to coordinate: a frontend and a backend agent agreeing on an API, two teammates' agents handing over context, or a friend's agent answering a question about their code. You stop copy-pasting between chat windows.
## Quick start
You need Node.js 22.18+ (`brew install node` or [nodejs.org](https://nodejs.org)) and git, on macOS or Linux (on Windows, use WSL).
**1. You:** install, then create an invite.
```bash
curl -fsSL https://raw.githubusercontent.com/pandeyvishwas51-oss/wokitoki/main/install.sh | bash -s -- --name "Vishwas"
wokitoki pair create
```
**2. Your friend:** paste the one line that `pair create` printed for them. It looks like this:
```bash
curl -fsSL https://raw.githubusercontent.com/pandeyvishwas51-oss/wokitoki/main/install.sh | bash -s -- --name "Rahul" --join amp2.XXXX...
```
That single command installs WokiToki, starts it at every login, connects it to their Claude Code and Codex, and joins your invite.
**3. Both of you:** compare the 6-digit verify code on a call. Approve only if the codes match.
```bash
wokitoki pair approve <request-id>
```
Done. `wokitoki agents ls` shows your friend as `active [online]`.
<p align="center"><img src="assets/how-it-works.png" width="900" alt="How WokiToki works: 1 install with one line, 2 pair by comparing a 6-digit code, 3 agents talk directly, encrypted end to end"></p>
## Use it from Claude Code or Codex
Restart any open agent sessions once after setup, then just ask:
> *Check my WokiToki inbox.*
>
> *Send Rahul a message on WokiToki asking whether the orders API returns amounts in paise.*
| MCP tool | What it does |
|---|---|
| `wokitoki_get_status` | your identity, network state and outbox |
| `wokitoki_list_peers` | paired agents, their state, online or offline |
| `wokitoki_send_message` | send encrypted text to a paired agent |
| `wokitoki_read_inbox` | read unread messages (marks them read) |
| `wokitoki_list_messages` | conversation history with delivery state |
Your agent **cannot** pair, approve or revoke; those always need you at the terminal. Incoming text is labelled as untrusted data, so agents treat it as information, never as instructions. Any MCP client can use WokiToki by running `wokitoki mcp`.
From the terminal:
```bash
wokitoki msg send Rahul --text "Can you confirm the /orders response shape?"
wokitoki msg watch # live incoming messages
wokitoki msg ls # history: queued -> relayed -> delivered
wokitoki agents revoke Rahul # cut them off, on both sides, immediately
```
## How it works
<p align="center"><img src="assets/p2p.png" width="900" alt="WokiToki peer-to-peer architecture: each daemon announces its public key on a public DHT, the peer looks it up, and the two connect directly with a Noise-encrypted channel through NAT hole punching. Nothing is stored in between."></p>
- Each WokiToki daemon joins the public **HyperDHT** (an open, distributed network, also used by the Keet messenger) under its own Ed25519 public key. Your invite carries that key.
- Your friend's daemon looks the key up. The two machines then **hole-punch** through their home or office routers and open a **Noise-encrypted, mutually authenticated** connection. WokiToki's own message encryption (XChaCha20-Poly1305) runs inside it.
- A firewall drops every key that is not a paired peer. A stranger can only knock while you have an invite open, and gets 5 tries.
- If your friend is offline, messages wait encrypted in your outbox and deliver automatically when both machines are online.
## FAQ
### How do I make two Claude Code instances on different computers talk to each other?
Install WokiToki on both computers, create an invite on one with `wokitoki pair create`, and run the printed one-liner on the other. Compare the 6-digit code and approve on both sides. Both Claude Code sessions then have `wokitoki_send_message` and `wokitoki_read_inbox` tools.
### Can Claude Code talk to OpenAI Codex?
Yes. WokiToki is a standard MCP server, so any mix works: Claude Code with Claude Code, Claude Code with Codex, or Codex with Codex. Each person keeps their own AI subscription and account.
### Do I need to host a server, open a port or set up a tunnel?
No. Connections are peer to peer over a public distributed hash table with NAT hole punching. There is no WokiToki server, no signup and no port forwarding. A self-hostable relay exists only as an optional fallback for networks that block peer-to-peer traffic.
### Is it secure? Can the other person's agent run commands on my machine?
Messages are end-to-end encrypted with libsodium, and pairing needs a one-time invite plus both people approving a matching verify code. WokiToki only carries text: it never runs incoming content, never touches your files, and gives the other side no access to your machine. Read the [threat model](docs/THREAT_MODEL.md) for the honest limits. There has been no external audit yet.
### What happens if my friend's laptop is off?
Your message waits, encrypted, in your local outbox and delivers when both computers are online at the same time. The daemon starts at login, so in practice that means "both laptops awake".
### Does it work behind corporate firewalls?
Usually, through NAT hole punching. Networks that block UDP entirely, or strict carrier NAT on both ends, can prevent a direct connection. Then use the optional relay: `wokitoki config set-relay <url>`.
### How is WokiToki different from Tailscale-based tools or Google's A2A protocol?
Tools built on Tailscale need every participant in the same tailnet with an account. A2A is an interoperability protocol for agents that expose HTTP endpoints. WokiToki needs neither an account nor a reachable endpoint: two developers pair with an invite and a verify code, and their existing coding agents talk through MCP.
### Is WokiToki free?
Yes. It is open source under Apache-2.0, and there is no paid tier, hosted service or telemetry.
## Compared with the alternatives
| | WokiToki | Copy-paste between chats | Shared Slack/Discord bot | Tailscale-based agent tools |
|---|---|---|---|---|
| Agents talk directly | yes | no | via a third-party server | yes |
| Server or account needed | none | none | bot hosting + workspace | Tailscale account and tailnet |
| End-to-end encrypted | yes | n/a | no | transport (WireGuard) |
| Pairing with human verification | 6-digit verify code | n/a | no | tailnet ACLs |
| Works with Claude Code and Codex via MCP | yes | manual | custom integration | varies |
## Security, honestly
- Messages are end-to-end encrypted between the two daemons. In peer-to-peer mode, nothing is stored anywhere but the two machines.
- Pairing needs a one-time invite **and** both people approving the same verify code. A leaked invite alone pairs no one, and nobody in the middle can swap keys without detection.
- Message-layer keys are static per pairing (no ratchet yet). Keys live in `~/.wokitoki/identity.json` (mode 0600), and local history is stored unencrypted on disk.
- No external security audit has been done. Don't send secrets.
Details: [threat model](docs/THREAT_MODEL.md) · [architecture](docs/ARCHITECTURE.md) · [protocol](docs/PROTOCOL.md) · [security policy](SECURITY.md) · [verification log](VERIFICATION.md)
## Commands
```
wokitoki setup [--name <n>] [--join <invite>] [--relay <url>] [--no-autostart] [--no-mcp]
wokitoki pair create [--expires 10m] | join <invite> | requests | approve <id> | reject <id>
wokitoki agents ls | revoke <agent>
wokitoki msg send <agent> --text <t> [--reply-to <id>] | ls [agent] | watch [agent]
wokitoki status | logs | identity show | mcp
wokitoki config show | set-relay <url|off>
wokitoki daemon run|start|stop|status|install|uninstall
```
Every command accepts `--json` and `--home <dir>` (default `~/.wokitoki`). The old command name `agentmesh` still works as an alias.
## Daily use
- **After a reboot:** nothing to do; the daemon starts at login. Check with `wokitoki status`.
- **Update:** rerun the install one-liner, then `wokitoki daemon stop && wokitoki daemon start`.
- **Uninstall:**
```bash
wokitoki daemon uninstall
rm -rf ~/.wokitoki ~/.wokitoki-src ~/.local/bin/wokitoki ~/.local/bin/agentmesh
claude mcp remove wokitoki -s user
codex mcp remove wokitoki
```
## Development
```bash
pnpm install
pnpm check # strict TypeScript + 57 tests: unit, adversarial, P2P on a local DHT testnet, multi-process CLI e2e, real MCP calls
```
[Roadmap](ROADMAP.md): tasks, encrypted file transfer, multi-agent rooms and a web console are next. License: [Apache-2.0](LICENSE).
<p align="center"><img src="assets/logo.svg" width="72" alt="WokiToki mascot"></p>
This server cannot be deployed
Maintenance
ActivityMaintained
ResponsivenessNo issues