hermesh

Your AI assistants hand coding work to Claude Code and Codex
on your own computers.

You approve each one on your own device. Open source, and you can run the server yourself.

Get started · Watch it in 50 seconds · How it works · Security · Docs

What it's for

You: From now on, when someone reports a bug in our repo, have Claude Code on my Mac mini fix it.

The next morning, a new issue: "Nothing happens on Sign in".

Grok Bot: Claude Code on your Mac mini fixed #142: the email wasn't checked before submitting. Here's PR #143.

Your agent starts a session on your machine, follows it, answers the coding agent's questions or passes them to you, and brings back the result and the diff. Your code, and your Claude Code or Codex sign-in, stay on your machine.

How it works

Three words to know:

A Mac can be your trusted device and a machine at once.

Get started

You need a trusted device, a machine and an agent. Start from what you have:

With a Mac

1. Download Hermesh for Apple silicon or Intel (macOS 14 or later), drag it to Applications and open it.
2. Click Get Started and sign in at hermesh.io in your browser, with an email code, Google or GitHub. The first sign-in creates your account.
3. This Mac is now your trusted device and your first machine. Add a folder your agents may work in as a project, then connect your agent.

The app keeps itself up to date.

With an iPhone and a server

1. Install Hermesh for iPhone (iOS 26 or later; coming to TestFlight soon) and tap Sign In. Its Get Started card leads you through the rest.
2. On the server, run:

   curl -fsSL https://github.com/hermeshai/hermesh/releases/latest/download/install.sh | sh
   

   Scan the code it shows with the app, approve with Face ID, and type on the server the two digits your phone shows. Run as root, it asks which user Hermesh runs as. From then on it starts with the server.
3. As that user, install Claude Code or Codex and sign in once (run claude); hermeshd doctor checks everything. Add a project from the app, then connect your agent.

Only Linux

On a Linux computer that runs no coding agents (your desktop, or a user of its own):

curl -fsSL https://github.com/hermeshai/hermesh/releases/latest/download/install.sh \
  | sh -s -- --device

With no other device yet, it signs you in with a code your browser shows, and this computer becomes your first trusted device, its key sealed by a passphrase you choose. Then add machines as above, and approve what asks to join with hermeshd approve.

Connect your agent

Choose Connect Agent… in the app (Connect an Agent… on the iPhone) and send your agent the line it gives you:

Connect to Hermesh: follow https://hermesh.io/connect

• Online assistants (ChatGPT, Claude.ai, Grok Bot, Muse) add the MCP server https://mcp.hermesh.io/mcp and send you to sign in.
• Agents on a computer (Claude Code, Codex, Cursor, your own) install the CLI and sign in:

  curl -fsSL https://github.com/hermeshai/hermesh/releases/latest/download/install.sh \
    | sh -s -- --cli
  hermesh auth login --name "Codex on MacBook"
  

Its request comes to your trusted device: choose the projects it may use, and approve. Then ask it: "Use Hermesh to have Claude Code say hello to me in the app project."

More machines

Run the same command on each one and scan its code (in the Mac app: Enter Pairing Code…). On a machine, hermeshd doctor says what to fix and hermeshd project add <folder> adds a project; docs/hermeshd.md has every command.

Works with

Agents that hand out the work: assistants that speak MCP, such as ChatGPT, Claude.ai, Grok Bot, Muse and Cursor, and any agent that can run a command line, through the hermesh CLI.

Coding agents that do it, on your machines: Claude Code and Codex, and through the Agent Client Protocol Gemini CLI, Cursor, GitHub Copilot CLI, OpenCode, Qwen Code, Goose, Augment, Kimi, Kiro, Qoder, Grok, DeepSeek Harness and CodeBuddy Code.

ACP agents run as one long-lived process each, with multi-turn conversation, interruption and resumption, and every tool call and permission request seen by Hermesh. Their sign-in can't be checked in advance: run each CLI once as the machine's user first. Whichever is installed (and allowed by the grant) shows up in hermesh project list.

Security

• You approve everything, on your own device. Adding a machine, an agent or a device takes Face ID, Touch ID or your passphrase. Your key stays on your trusted devices (in the Secure Enclave on a Mac or an iPhone), so neither hermesh.io nor your own server can approve, or forge a command.

• Encrypted end to end, with one difference:

  Your agent connects with	Who can read the content
  The hermesh CLI	Only your agent and your machine: hermesh.io, or your server, passes on encrypted bytes
  MCP (online assistants)	Also the MCP server, which holds that connection's key (Privacy Policy). Run your own server to keep it on yours

• Each agent gets a grant: which machines, projects and coding agents, and for how long. Revoke it at any moment.

• A grant is full trust. There, a coding agent can do anything you can: read credentials, use the network, push code. Hermesh adds no sandbox. Grant access only to agents you trust; if one reads untrusted web pages or email, a prompt injection reaches as far as its grant.

• Your own work stays apart. A session works in its own git worktree, on a hermesh/<session> branch, and never pushes for you: review the diff before merging. An in-place project lets the agent write straight into the folder.

• The center is trusted once, at most. An agent pairing through hermesh.io takes your account's owner on the center's word, once; your machines accept only grants your key signed. Pair with your own server's connection address to trust no center at all.

• Lose a device? Remove it from another trusted device. With only one you'd set up again, so add a second: another iPhone, a Mac or a Linux computer.

More in the technical design, trusted devices and the MCP gateway.

Run your own server

Instead of hermesh.io, run the relay on a server with a domain name:

docker run -d --name hermesh-relay -p 443:443 -p 80:80 -v hermesh-relay:/data \
  ghcr.io/hermeshai/hermesh-relay:latest serve --data-dir /data --domain relay.example.com --mcp
docker logs hermesh-relay   # shows the claim code

In the Mac or iPhone app, choose Use Your Own Server… and enter the server's address and the claim code. Machines and agents then use the connection address the app gives you:

curl -fsSL https://github.com/hermeshai/hermesh/releases/latest/download/install.sh \
  | sh -s -- relay.example.com/o_xxxxxxxx
hermesh auth login relay.example.com/o_xxxxxxxx

--mcp also serves MCP at https://relay.example.com/mcp; leave it out if your agents all use the CLI. Don't run the relay as the same user as a machine on that server.

Questions

No. The iPhone app or a Linux computer can be your trusted device, and your machines can be Linux servers.

Not what goes through the CLI: that is end to end. Over MCP, its MCP server holds the connection's key, so it can read what that assistant sends and receives. Your own server keeps that to you.

No. They run on your machine with their own sign-in; Hermesh starts them there and never reads their keys.

macOS, or Linux on x86-64 or ARM64; on Linux, systemd keeps Hermesh running (elsewhere, run hermeshd run yourself). And a coding agent, signed in as the same user.

Documentation

Components

Development

make generate && make lint test   # code from proto, static checks, unit tests
make e2e                          # full user flows with real binaries and scripted agents

More, and how to release: docs/development.md.

License

Apache License 2.0, see LICENSE.