Covalent Bond

[Covalent Bond: two agents sharing an encrypted channel]

Covalent Bond

A secure channel that lets two AI coding agents on different machines work as one.

Quick start · How it works · Tools · Security · Architecture · Contributing

[MCP compatible] [Node.js >= 18] [E2EE: AES-256-GCM + X25519] [Relay: Cloudflare Workers] [License: MIT]

A covalent bond is two atoms sharing a pair of electrons. Covalent Bond is two coding agents on different machines sharing something just as fundamental: a private, authenticated channel, and through it, a way of working.

Picture two developers pairing across a distance. One has their agent dialed in (organized skills, sharp conventions, the right context loaded) and it produces excellent work. The other is more ad-hoc, and their agent turns out something that's just okay. Covalent Bond is the channel that closes that gap: the well-tuned side can hand over the files, context, and conventions that make its agent good, so both agents converge on the same high-quality way of working. It bonds not just two agents, but two ways of working, pulling the weaker setup up to the level of the stronger, as equal peers.

It's peer-to-peer and end-to-end encrypted: the two agents pair up and exchange messages and files through a relay that can never read the traffic. It works with any MCP-compatible agent (Claude Code, Cursor, Codex, Windsurf, Cline, and others), exposing bond_* tools over the Model Context Protocol. And it keeps a human in the loop: every incoming file waits for explicit consent before it touches disk.

• 🤝 Bonds two ways of working: share files, context, and conventions so both agents level up together
• 🔒 Authenticated E2EE: a malicious relay can't read or MITM the channel
• 🌐 Cross-machine: pair agents anywhere, through a dumb relay (no inbound ports)
• 🙋 Human-in-the-loop: consent before any file is written; full audit log
• 🔌 Agent-agnostic: any MCP client, not tied to one vendor

It doesn't magically clone one setup onto another. It's the secure, consent-gated channel that makes deliberately sharing your working style possible between two machines that otherwise can't reach each other.

Quick start

Shortcut: let your agent set everything up. Paste docs/AGENT-SETUP.md to any coding agent and say "set up Covalent Bond". It walks the agent through install, relay config, MCP registration, and pairing, with clear stops for the parts only a human may do.

Or do it yourself on both machines (steps 1–2), then pair (step 3).

1. Register the MCP server with your agent — no install needed, npx fetches the npm package. For Claude Code:

   bash

   claude mcp add --scope user --env COVALENT_RELAY_URL=https://covalent-bond-relay.gopalrajsuresh.workers.dev covalent -- npx -y covalent-bond

   (--scope user makes the bond_* tools available in every project. The env var points at the public relay; to self-host instead, see Pick a relay.)

   For other MCP clients (Cursor, Codex, Windsurf, …), this standard stdio-server config works in most of them; paste it into the client's MCP settings JSON (the top-level key is usually mcpServers or servers):

   json

   {  "mcpServers": {    "covalent": {      "command": "npx",      "args": ["-y", "covalent-bond"],      "env": {        "COVALENT_RELAY_URL": "https://covalent-bond-relay.gopalrajsuresh.workers.dev"      }    }  }}

   Windows and local-testing notes

   On Windows, if your client struggles to spawn npx directly, wrap it: "command": "cmd", "args": ["/c", "npx", "-y", "covalent-bond"]. For local testing without a deployed relay, run the mock relay from a source checkout (npm run relay:dev) and point COVALENT_RELAY_URL at http://localhost:8787.

   Prefer running from source?

   bash

   git clone https://github.com/gopalrajsuresh/covalent-bond.git covalent-bondcd covalent-bondnpm installnpm test                    # optional but recommended; all suites run offlinecp .env.example .env        # PowerShell: Copy-Item .env.example .env

   The .env already contains the public relay URL. Then register with command: node, args /absolute/path/to/covalent-bond/bin/cli.js (Claude Code: claude mcp add --scope user covalent -- node /absolute/path/to/covalent-bond/bin/cli.js), no env var needed.

2. Restart your agent session, then verify: ask the agent to run bond_status. It should reply "No active session", which means the tools are live.

3. Pair and share

   • Machine A: "Create a Covalent Bond session" → agent runs bond_connect → you get a code like AbCd-1234-XyZw.
   • Tell the code to the other human out-of-band (chat or voice, never through the relay).
   • Machine B: "Join Covalent Bond session AbCd-1234-XyZw" → agent runs bond_join.
   • Either side: bond_status → Secure channel established → send files (bond_send) and messages (bond_message). Incoming files wait for the human to bond_accept.
   • Done? bond_end on both sides.

How it works

Both agents meet at a relay that is untrusted by design — it forwards ciphertext it can never decrypt. The session code travels human-to-human, never through the relay, and that out-of-band secret is what authenticates the channel:

sequenceDiagram    participant A as Agent A (host)    participant H as Humans (out-of-band)    participant R as Relay (untrusted)    participant B as Agent B (guest)
    A->>A: bond_connect → session code    A-->>H: code shared by chat / voice    H-->>B: code entered on the other side    B->>B: bond_join(code)    A->>R: ephemeral public key (routing ID only)    B->>R: ephemeral public key (routing ID only)    Note over A,B: key = X25519 ⊕ session code → AES-256-GCM    A->>R: ciphertext    R->>B: ciphertext (relay sees nothing else)    B->>B: file held until human runs bond_accept

If anything in the middle tampers with the key exchange, key confirmation fails and the session aborts before any data flows — a malicious relay can't read or impersonate either side.

Pick a relay

The relay is a dumb pipe and untrusted by design: it sees only a random routing ID and ciphertext, never the session code, keys, or plaintext, and a malicious relay still can't read or forge anything (that's covered by the test suite's MITM scenario). So either option below is equally secure; it's purely a convenience choice.

Option 1: use the public relay (fastest)

COVALENT_RELAY_URL=https://covalent-bond-relay.gopalrajsuresh.workers.dev

Sessions expire after 30 minutes of inactivity and nothing is retained. Best-effort availability, rate-limited.

Option 2: deploy your own in ~2 minutes (free Cloudflare account)

npm install -g wranglerwrangler login          # opens the browser; create a free account if you don't have onecd cloudflare-workerwrangler deploy         # prints your relay URL: https://covalent-bond-relay.<your-subdomain>.workers.dev

Then set COVALENT_RELAY_URL to that URL on both machines. Full details (Durable Object migration, optional per-IP throttle, local wrangler dev) are in cloudflare-worker/README.md.

Pair two agents

1. Machine A: "Create a Covalent Bond session." → the agent calls bond_connect and returns a code like AbCd-1234-XyZw.
2. Share that code with Machine B out-of-band (chat, voice, in person). It's the secret that secures the link. Never paste it into the relay.
3. Machine B: "Join Covalent Bond session AbCd-1234-XyZw." → the agent calls bond_join.
4. When bond_status shows Secure channel established, start sharing: "Send src/auth.js to my peer," or hand over the pieces that make your agent good: a skill file, a conventions doc, the context that shapes how it works. The peer sees a consent prompt and accepts before anything is written, and the received content arrives wrapped as untrusted data for the other agent to read and adopt.

Tools

• bond_connect — Create a session and get a shareable session code
• bond_join — Join a session using a code shared out-of-band
• bond_status — Connection status, pending transfers, and new events
• bond_send — Send a file to the connected peer, encrypted end to end
• bond_message — Send a short encrypted text message to the peer
• bond_wait — Long-poll for the next peer event
• bond_accept — Accept a pending file transfer after human consent
• bond_decline — Decline a pending file transfer
• bond_end — End the session and disconnect from the relay

Full reference:

? marks an optional parameter. confirmed channel means the tool refuses to run until key confirmation has succeeded on both sides; human consent means the agent may only call it after the human explicitly decides on the pending transfer.

Incoming events raise a desktop notification (disable with COVALENT_NOTIFICATIONS=off). File size defaults to 256 KB, overridable via COVALENT_MAX_FILE_KB (up to 384 KB, bounded by the relay's wire limit).

Security

Each session is secured by two secrets: ephemeral X25519 keys exchanged through the relay, and a short session code the two humans share out-of-band. The code is mixed into the encryption key but never sent to the relay, so a relay that tampers with the key exchange (a MITM) can't derive the key, key confirmation fails, and the session aborts before any data flows. The relay only ever sees a routing ID, public keys, and AES-256-GCM ciphertext. Every incoming file waits for explicit human consent, and every operation is logged.

Full details, threat model, and handshake diagram: docs/ARCHITECTURE.md.

Share the session code over a channel the relay operator doesn't control. That out-of-band step is what makes the channel authenticated, not just encrypted.

Testing

npm test          # all suites, via test/run-all.js, against an in-process mock relay

No network or wrangler needed. The suite covers the key schedule, the full handshake and file transfer, two MCP servers end-to-end, a simulated malicious relay (mitm), and the relay's hardening. See docs/ARCHITECTURE.md for the breakdown, and docs/TWO-MACHINE-TEST.md for a real two-machine run.

Deploying the relay

The relay is a Cloudflare Worker in cloudflare-worker/ with one Durable Object per session. It stores only routing IDs, public keys, and ciphertext, expiring 30 minutes after the last activity. It can't decrypt anything.

Status

Authenticated E2EE, MCP integration wired end-to-end, full test suite green. Pre-1.0: the protocol and API may still change.

Contributing & security reports

Contributions are welcome. Start with CONTRIBUTING.md, which includes the security invariants and the branch/PR workflow. To report a vulnerability, please use the private process in SECURITY.md, not a public issue.

License

MIT; see LICENSE.

_{Two machines. One way of working. Back to top ↑}