Another
Several agents, one project, one machine.
Another is a background messaging system that helps coding agents work on the same project on the same machine at the same time. It helps reduce editing conflicts, exchange updates and reuse registered resources. Free and open source, always. Measured resource use records a sample on one Mac and its limits.
It also works across machines today, through a self-hosted federation hub or authenticated remote peers. The hub can read mail it relays; remote peers cannot reap processes. See federation and remote access.
Claude Code and Codex cannot see each other without something shared. They can edit the same files, start the same fix twice, or launch duplicate model servers. Another combines path holds, messages and requests, activity, and a registry for ports and model servers. Its mailbox is a way to address agents and exchange messages. It does not launch agents, assign their work, or enforce a resource limit.
Home: toanother.one · License: Apache-2.0 · Status: Milestones 1-6 complete. Hosted connectors (OAuth for ChatGPT, Claude, and other MCP clients) and the loopback dashboard are in this tree. See docs/REMOTE.md and docs/DASHBOARD.md.
Why
This is the failure I kept hitting:
- Two agents write the same file and the last save wins. Git shows the wreck. It does not say which session was in the file.
- Neither session can see the other, so both implement the fix.
- The handoff is me, copying a paragraph from one terminal to the other.
- A second agent starts another model server, or binds a port the first one already has.
What they get instead:
- Path holds, taken at the edit. With the hook pack, each file an agent edits gets an expiring hold, and a Claude Code, Codex or Grok CLI edit to a file another agent holds is stopped with the holder's name, across worktrees and clones of the same repo.
another hold check exits 3 on overlap and grants nothing. Edits made from a shell command get past the hook, and Cursor has no edit guard, so use another checkout when you need a real boundary.
- Mail and requests.
another send --request opens work for an agent to claim. Claiming assigns an owner. An ack records receipt. If the request goes stale, someone else can take it.
- An activity feed, so a session can ask who is doing what before it starts a duplicate.
- A resource registry, so an agent borrows a running server instead of launching a duplicate.
another reap needs a human --confirm on that machine. Remote peers cannot reap.
Hooks put a short digest in at session start, and a one-line notice after the next tool call when mail lands mid-task. That is why the agents look.
State is one SQLite database on the machine (~/.another, or $ANOTHER_HOME). A second machine can federate through a hub you host. The hub can read mail it relays.
OAuth for hosted MCP clients (ChatGPT, Claude.ai) is in this tree, and so is the loopback dashboard (another dashboard, or /ui on a loopback another serve). See the changelog and how I run the agents on this repo.
The website source lives in site/. sh site/build.sh checks the pages, and python3 scripts/check-copy.py rejects em dashes and a banned-phrase list.
Quickstart
1. Install with Rust 1.88 or newer:
cargo install --locked --git https://github.com/tamler/another another
Or build from a checkout (the binary ends up at target/release/another):
git clone https://github.com/tamler/another && cd another && cargo build --release --locked
Once v0.1.0 is published, prebuilt binaries for macOS (arm64, x86_64) and Linux (static, x86_64, arm64) will be on GitHub Releases: unpack one and put another on your PATH. On macOS a downloaded binary may need its quarantine flag removed: xattr -d com.apple.quarantine ./another.
以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗
