Intercom

Claude Intercom

Two-way communication bridge between Claude Code sessions using the Channels API.

Let two Claude Code instances on different machines talk to each other in real-time. One sends a message, the other receives it instantly as a channel notification and can reply back.

What Intercom does that native can't

Claude Code has native cross-session messaging now. Here's where Intercom is still the answer:

• Two different people. This is the big one. Native connects your sessions — the inbox is tied to your own OS user and cross-machine delivery rides your own Remote Control. A backend dev on their laptop and a frontend dev on theirs are two accounts, and native can't join them. That two-person hotline is exactly what Intercom was built for.
• Native Windows. Native messaging is macOS and Linux only. Intercom is Bun over HTTP and doesn't care what OS you run.
• Your network, not Anthropic's. Native cross-machine messages relay through Anthropic servers. Intercom POSTs straight over your own network, VPN or tunnel.
• Any provider. Works on Bedrock, Vertex and Foundry, and in telemetry-disabled environments, where native messaging is switched off.
• Container to host. Native discovery works through files on disk, so a session in a container and one on the host can't see each other. Intercom just needs a reachable port.
• Honest delivery. send_message reports sent, delivered-to-process and answered as three separate states, plus a check_message tool. You find out when the other side never picked it up, instead of planning around an answer that was never coming.

If you're on macOS or Linux and you just want your own sessions talking to each other, native is simpler — use that. Everything else above is what this is for.

Why?

If you have a backend developer and a frontend developer each running Claude Code on separate machines, they currently have to relay questions through Slack/Discord/copy-paste. Claude Intercom creates a direct hotline between the two AI sessions — one Claude can ask the other about endpoints, schemas, or implementation details and get answers from the actual codebase.

Architecture

Machine A                          Machine B(e.g. backend dev)                 (e.g. frontend dev)
Claude Code A                      Claude Code B     |                                  |     +-- intercom.ts ---HTTP POST----> intercom.ts --+     |   (channel)     <--HTTP POST--  (channel)     |     |                                               |     +-- stdio (MCP) --+        +-- stdio (MCP) -----+                        |        |                   Claude A    Claude B

Both instances run the same intercom.ts file. Each listens for HTTP messages and pushes them into its local Claude Code session via the Channels API. Each also exposes send_message and check_message tools that Claude can call.

Requirements

• Claude Code v2.1.80 or later
• Bun runtime
• Both machines must be able to reach each other over HTTP. Tailscale is the recommended way — see Connecting the two machines

Set it up with Claude

Paste this into a Claude Code session on each machine and it will walk you through the whole thing. Run it on machine A first, keep the secret it gives you, then run it on machine B.

Set up claude-intercom on this machine so this Claude Code session can messagea Claude Code session on my other machine.
Work through this in order, and stop and ask me whenever you need somethingonly I can tell you.
1. Check prerequisites. Confirm `bun --version` works and that `claude --version`   is 2.1.80 or newer. If Bun is missing, tell me how to install it for my OS   and stop there.
2. Work out how the two machines will reach each other. Run `tailscale ip -4`.   If that returns an address we'll use Tailscale. If Tailscale isn't installed,   tell me it's the recommended option and ask whether I want to install it or   use something else.
3. Ask me these, one at a time:   - a short role name for THIS machine, e.g. "frontend", "vps", "laptop"   - the address of the OTHER machine (hostname or IP)   - whether this is the first machine I'm setting up or the second
4. Handle the shared secret.   - First machine: generate a strong random secret, show it to me once, and     tell me I'll need it when I run this on the other machine.   - Second machine: ask me to paste the secret from the first one.   Both machines must end up with exactly the same secret.
5. Write `.mcp.json` in this project with an "intercom" server: command `npx`,   args `["-y", "claude-intercom"]`, and env `MY_ROLE`, `REMOTE_HOST` (the other   machine plus `:8788`), `INTERCOM_SECRET`, `INTERCOM_PORT` set to 8788.   If `.mcp.json` already exists, merge the intercom entry into it instead of   overwriting the file. Make sure `.mcp.json` is gitignored, it holds the secret.
6. Tell me to restart Claude Code on this machine with:   `claude --dangerously-load-development-channels server:intercom`   and explain that the flag is needed because the channels capability is still   experimental, and that messages only arrive while both sides are running.
7. Once both machines are up, tell me to test it by asking one session to send   a message to the other. Remind me that `send_message` reports sent /   delivered-to-process / answered separately, so "delivered" means the remote   process accepted it, not that the other Claude has read it.
Don't invent configuration keys. The only environment variables are MY_ROLE,REMOTE_HOST, INTERCOM_SECRET, INTERCOM_PORT and INTERCOM_SEND_TIMEOUT_MS.

Prefer to do it by hand? The manual steps are below.

Quick Start

1. Install

Nothing to install. npx fetches it on first run, on both machines.

You do need Bun on your PATH — the server uses Bun.serve for its HTTP listener, and the npx entry point hands off to it.

git clone https://github.com/MuhammadTalhaMT/claude-intercom.gitcd claude-intercombun install

2. Configure

Copy the example config into your project's .mcp.json.

First generate a secret, and use the same value on both machines:

openssl rand -base64 32

Then export it in the shell you launch Claude Code from, rather than typing it into the config:

export INTERCOM_SECRET='the-value-you-just-generated'

.mcp.json is a project file, and project files get committed. Claude Code expands ${VAR} in .mcp.json — in command, args, env, url and headers — so the config below can be checked in and shared with your teammate while the secret itself never leaves your environment.

Machine A (e.g. backend — static IP or VPS):

{  "mcpServers": {    "intercom": {      "command": "npx",      "args": ["-y", "claude-intercom"],      "env": {        "MY_ROLE": "backend",        "REMOTE_HOST": "MACHINE_B_IP:8788",        "INTERCOM_SECRET": "${INTERCOM_SECRET}",        "INTERCOM_PORT": "8788"      }    }  }}

Machine B (e.g. frontend — can be behind NAT):

{  "mcpServers": {    "intercom": {      "command": "npx",      "args": ["-y", "claude-intercom"],      "env": {        "MY_ROLE": "frontend",        "REMOTE_HOST": "MACHINE_A_IP:8788",        "INTERCOM_SECRET": "${INTERCOM_SECRET}",        "INTERCOM_PORT": "8788"      }    }  }}

3. Start

On both machines:

claude --dangerously-load-development-channels server:intercom

4. Talk

In either Claude Code session:

"Send a message to the other developer asking what API endpoints are available for the dashboard."

Claude will use the send_message tool to POST the message to the other machine. The other Claude receives it as a channel notification and responds.

Connecting the two machines

Use Tailscale. Install it on both machines and they get stable private addresses on your own tailnet. Then point each side at the other's tailnet address:

"REMOTE_HOST": "other-machine:8788"

This is strictly better than exposing a port to the internet: no public listener, no port forwarding, the address doesn't change when your ISP reassigns your IP, and device identity is enforced by Tailscale rather than resting entirely on a shared string. Set INTERCOM_HOST to your tailnet address — or to 127.0.0.1 if you are also fronting it with a tunnel — to be certain nothing outside can reach it at all.

# On the machine behind NATngrok http 8788

"REMOTE_HOST": "your-subdomain.ngrok-free.app"

Tools

send_message

Sends a message and returns immediately with an id. It does not wait for an answer — the reply arrives later as its own inbound message.

The acknowledgement deliberately does not claim the other developer received it. An HTTP 200 proves the remote process accepted the message, not that the other Claude ever read it.

check_message

Answers the question a fire-and-forget channel otherwise can't: is this reply slow, or is it never coming? Pass an id, or omit it to list everything outstanding.

Every message sits in one of three states:

expectReplyWithin is what makes "overdue" mean anything, and it's per message rather than one global timeout — so a quick endpoint lookup and a long investigation aren't judged on the same clock.

Configuration

How It Works

1. Claude Code spawns intercom.ts as a subprocess (MCP server over stdio)
2. The script declares claude/channel capability — this registers it as a Channel
3. It starts an HTTP server listening for incoming messages
4. When a message arrives (authenticated via shared secret), it calls mcp.notification() with method: 'notifications/claude/channel'
5. Claude Code surfaces the notification in the conversation as a <channel> tag, carrying the message id
6. Claude reads it and can reply using send_message with replyTo set to that id, which POSTs to the remote machine
7. The original sender matches the replyTo against its own outbound record and marks that message answered

The stdio leg is not an implementation detail you can swap for HTTP. It is what attaches the server to a specific live Claude Code session — the claude/channel capability is registered over that connection, and it's the reason mcp.notification() lands in a conversation at all. Host this remotely and the notification goes to whatever MCP client connected instead.

Security

• Shared secret authentication: POST /message requires an X-Intercom-Secret header matching the configured secret, compared in constant time. Anything else gets a 401 Unauthorized. GET /health is deliberately not authenticated, so you can verify a tunnel end to end — it reports this instance's role and version to anyone who asks, so treat a reachable intercom as discoverable.
• The secret is only as private as the transport: it is sent as a plaintext header on every message. Over Tailscale (WireGuard) or an HTTPS tunnel that is fine. Over plain HTTP on a shared network, anyone on the path can read it and then use it.
• No replay protection: messages carry an id and a timestamp, but neither is checked for freshness or reuse. Someone who captures a single authenticated request on a cleartext link can resend it verbatim, as often as they like.
• No default secret: INTERCOM_SECRET has no fallback value. Leave it unset, leave a docs placeholder in place, or reference a ${VAR} you never exported, and the intercom starts unpaired — it binds no port and send_message refuses, explaining why. The unexpanded-${VAR} case matters because Claude Code passes a missing variable through as literal text, which would otherwise give both machines the same guessable secret.
• Inbound is treated as data, not instructions: the server tells the receiving Claude that a channel message comes from another person's session — it can't approve anything, can't change configuration, a slash command in the text is inert, and requests for credentials or env files should be refused and surfaced to you.
• No data persistence: Messages are forwarded in real-time and not stored.
• Configurable bind address: By default listens on 0.0.0.0 for cross-machine access. Set INTERCOM_HOST=127.0.0.1 when a tunnel is doing the reaching, so only the tunnel can connect.

Warning: Those instructions are a default, not a boundary. You cannot fix prompt injection with prompt instructions — anyone holding your secret and address can put text into your Claude session, and the only real limits are that session's own permission prompts. Native cross-session messaging enforces this properly with hold/accept/refuse inbound controls; this does not. Use a strong secret, keep it off the public internet, and don't pair with a peer you wouldn't hand a terminal to.

Endpoints

POST /message body:

{  "id": "8649f5da",  "replyTo": "40b0fd17",  "content": "3 routes, all behind requireTenant()",  "role": "backend-dev",  "timestamp": "2026-08-16T09:41:00.000Z"}

id and replyTo are optional — a message without them still delivers, it just can't be correlated.

The body is schema-validated before anything reaches your session. content is required and capped at 32,000 characters; id and replyTo must look like machine ids ([A-Za-z0-9_-], ≤64); role and timestamp may not contain < > " ' or any Unicode control, zero-width, or bidi-override character, because those are what a sender would use to forge the <channel …> wrapper the message is rendered inside. Anything else gets a 400, and a body over 64 KB gets a 413.

Use Cases

• Backend + Frontend collaboration: Backend Claude answers API questions from frontend Claude using the actual codebase
• Monorepo with split teams: Different Claude sessions working on different packages can coordinate
• Code review relay: One Claude reviews code and sends findings to the author's Claude session
• CI/CD notifications: Point your CI webhook at the intercom to push build results into a Claude session

Star History

License

MIT