agora-mesh

agent
Security Audit
Pass
Health Pass
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 34 GitHub stars
Code Pass
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Connect coding agents across your machines so they plan together, split the work, and push to separate git branches. Independent community project.

README.md

agora-mesh

Connect coding agents across your machines so they plan together, split the work, and push to separate git branches.

You have a laptop and a desktop, each with a coding agent. Today you copy context between them and decide who does what. With agora-mesh you give one task to one of them. The agents brief each other, challenge the plan, divide the work, and each pushes its part to its own branch. You review and merge.

Two machines, a laptop running Claude Code and a desktop running Codex, each with its own login and its own git clone. They plan and critique over an encrypted link. Each builds its part on its own branch in your git repository, and you review and merge.

  • Works with the agents you already use. Built for Claude Code. Codex, OpenCode and Hermes Agent have joined real rooms too.
  • Every machine keeps its own login and files. agora-mesh never sees a login or a key.
  • Nothing to host. No VPN, no server of your own, no MCP. Machines connect directly with end-to-end encryption. Public iroh relays help them find each other, and carry the traffic if no direct path exists.

Early development, version 0.6.0. An independent community project, not affiliated with Anthropic. What has and has not been tested.

Try it in one minute

You need one machine, git and an internet connection. No account, no API key, no Claude Code.

# macOS on Apple Silicon
curl -fLo agora https://github.com/kn0sti/agora-mesh/releases/latest/download/agora-aarch64-apple-darwin

# Linux on x64
curl -fLo agora https://github.com/kn0sti/agora-mesh/releases/latest/download/agora-x86_64-unknown-linux-musl

chmod +x agora
./agora demo

Other platforms are on the releases page.

Terminal recording of agora demo: two local peers connect after an approval, brief each other, one criticises the proposed split, they agree, each pushes its own branch, and the two branches merge without a conflict.

The connection, the room and the git branches are real, but the agents are scripted, so it shows the process and not the quality of a model. It uses no tokens and cleans up after itself.

Use it on your own project

claude plugin install agora-mesh --marketplace kn0sti/agora-mesh

The plugin adds the agora command inside Claude Code, so run the next steps there or ask your agent to. In a plain terminal, use the binary from above. INSTALL.md has the full steps, and cargo build --release builds it from source.

Connect two machines. Only connect machines you trust: a connected machine can run commands on yours.

agora setup --name MacBook     # once per machine
agora open                     # machine A: prints a network ID
agora join <network-id>        # machine B
agora status                   # both: compare the four words
agora approve <name>           # machine A

Start your first room. Each machine needs a clone of the same repository.

# machine B, once: register its clone
agora project add ~/code/my-app

# machine A: open the room
cd ~/code/my-app
agora discuss "Add rate limiting to the API" --approve

# machine B: accept the invitation
agora rooms
agora accept <room>

With --approve the room shows you the agreed split and waits. agora go <room> starts the work.

Use a clean clone. The agents work directly in it and switch its branch, so do not point them at the folder you are editing in.

When it helps

  • The task is too big for one session. Each agent has its own context and works on its own part at the same time.
  • The knowledge is on different machines. Each agent can use the files, tools and access of its own machine.
  • You want a second opinion. Put two strong models on one plan, or a different agent next to yours, and let them argue before anything is built.
  • You want one memory everywhere. Notes an agent saves appear on all your machines as Markdown, and can sync with Obsidian.

When it does not help: small tasks. There, one agent is faster and much cheaper.

What it costs

A room does not save tokens. Every agent re-reads the conversation on every turn. On one small planning question, a room of two used about 9 times the tokens of one agent, and a room of three about 13 times.

What one action costs, in median tokens: agora ask 1.8k, agora review 4.4k, one agent alone 4.5k, a room of two 41.6k, a room of three 60.5k.

Action Tokens (median) Time (median)
agora ask 1.8k 8 s
agora review, a small diff 4.4k 10 s
One agent alone 4.5k 68 s
Room of 2 41.6k 182 s
Room of 3 60.5k 272 s

So use ask and review for small things, and a room when the task is large enough to split or a second opinion is worth the price. These are repeated real runs on one cheap model, the rooms and the single agent without tools. Claude Code with its built-in tools starts each turn much higher, and the runs compare cost, not the quality of the plans. The full measurement, with an animation of a room spending tokens.

What happens in a room

The five steps of a room: brief, propose, critique, agree or ask you, build. The critique can send the plan back for revision. Every room has a budget, and if the agents cannot agree it stops and tells you where they differ.

In a real test run (three daemons on one Mac), three different agents exchanged 13 messages in 325 s and pushed one branch each, leaving main untouched (full log).

Animation of a real room, message by message. Three peers, one running Claude Code on a DeepSeek model, one running OpenCode on a DeepSeek model and one running Codex, exchange thirteen messages: a brief, what the other two know and lack, two answers, a pass, a split of the work, two agreements, the start of the work phase, and three reports that the work is done. At the end each peer has pushed its own branch and main is untouched.

agora-mesh or Hermes Agent?

They do different jobs.

Hermes Agent by Nous Research is a personal agent that grows with you: it builds memory and skills, runs scheduled jobs, and talks to you in Telegram, Discord or Slack. If you want one assistant that knows you, use Hermes.

agora-mesh does not replace that. It connects existing agents across machines so they can work on one task as equals.

agora-mesh Hermes Agent
What it is a link between separate agents one persistent personal and coding agent
Parallel work agents on different machines, one branch each helper agents under one parent, or a task board across profiles
Several machines each runs its own agent, login and clone one agent can run commands remotely over SSH or in sandboxes; its task board is single-host by design
Memory and skills that grow a shared notebook, no learning its main strength
Chat apps, scheduled jobs no yes

You can also use both. Hermes speaks the Agent Client Protocol, so it can join a room as a peer. In a real test, Claude Code and Hermes reached consensus on a design question in 8 messages, 135 s and 24,439 tokens, both on a DeepSeek model. Hermes only discussed, with its tools turned off; a work phase with Hermes has not been run (how to set it up). Facts about Hermes are from its documentation on 2026-11-21.

Security

A connected machine can make yours run commands. Agents need that to work together. This applies to whoever controls the other machine and to anything their agent reads, including web pages or files with a prompt injection. Connect only machines and people you would trust with a shell on your computer.

You decide how much runs without you. Agents discuss, plan and split. Two optional gates follow: the opener approves with discuss --approve, and a receiving machine approves with trust --work ask. Then the agents build on their branches. Without the gates a room that agrees starts working right away. agora-mesh does not merge anything.

Command What it does
agora discuss … --approve wait for your go before any work starts
agora trust <name> --work ask work from this peer waits for your go on this machine
agora trust <name> --rights read this peer's sessions here may only read (talk: no tools at all)
agora revoke <name> remove a peer
agora panic stop every room and session on this machine

By default a room that agrees starts working at once, with full rights. The details.

What has been tested

Ran for real: rooms and work phases between a Mac and a Linux machine with Claude Code; rooms and work phases with Codex, OpenCode and Claude Code on DeepSeek and GLM; a discussion with Hermes.

Never run so far:

  • two different accounts in one room
  • a real NAT between two networks
  • Windows, apart from an automated demo run
  • a real usage limit (the handover was simulated)
  • a work phase with Hermes, or Qwen Code in a room
  • agents other than OpenCode, Codex, Qwen Code and Hermes
  • larger tasks on models other than Claude

If you try one of these, report how it went.

Everyday commands

Command What it does
agora peers who is online and what they are doing
agora ask "<question>" ask another machine's agent one question
agora delegate "<task>" hand a task to another machine
agora review have another agent review your changes
agora discuss "<topic>" open a room: plan, split and build together
agora watch <room> follow a room live

All commands.

Learn more

License

MIT

Reviews (0)

No results found