Skip to main content
Glama

ReevesAgents is a free, open-source workspace for AI coding agents. It runs Claude Code, Codex, Kimi, OpenCode, Gemini CLI and 20 other coding CLIs side by side, each in its own tmux window, and gives you a terminal UI, a browser UI and a scriptable CLI over the same runs. Attach its MCP server to one CLI and that agent can spawn, read, steer and stop the others. Every CLI keeps its own login, and ReevesAgents keeps its state in a few JSON files under ~/.reeves.


Why ReevesAgents

You already use these CLIs, and you are already signed in to them. ReevesAgents runs them as a team on your machine and leaves every login where it was.

25 CLIs, one workspace

Claude Code, Codex, Kimi, OpenCode, Hermes, Gemini CLI and 19 more start from one command. Each runs as the real CLI in its own tmux window, so agents keep running after you close the UI.

One agent drives the rest

reevesagents attach cc gives Claude Code the MCP tools to spawn, read, steer and stop other agents. Codex, Kimi Code, Qwen Code, Hermes and OpenCode can host it too, and it stays off until you attach it (details).

Your logins stay put

Each CLI signs in the way it always does. ReevesAgents stores no API keys, adds no proxy, and redacts token-like text in tasks, notes and approvals before it writes state.

Four ways in

A terminal UI, a Web UI with a live pane per agent, a CLI with --json output for scripts, and the MCP. All four read and write the same runs.

Plain local state

A few JSON files under ~/.reeves, with no database and no Docker. The Web UI answers on 127.0.0.1 only and rejects foreign Host and Origin headers (guards).

macOS, Linux, WSL, Windows

tmux 3.0 or newer on macOS, Linux and WSL. Native Windows 10 1809 or later runs on the built-in winmux multiplexer over ConPTY, without WSL (2.1.0).

Related MCP server: openai-agents-mcp

A quick tour

Features

Run a team

  • reevesagents spawn provider[:nickname[:model]] starts a run. The first agent leads and the rest join as workers, and reevesagents add grows the latest run.

  • Per agent you choose the model, the permission mode (ask first, or skip with each CLI's own bypass flag), reasoning effort for Claude Code and Codex, and extra launch flags.

  • Presets save a live run and start it again later. Ended runs move to history, and reap ends zombie agents and orphan tmux sessions.

Watch and steer

  • peek reads recent output, send pastes text and key <id> enter submits it, interrupt sends Ctrl-C and open jumps to the agent's tmux window. stop and kill need --yes.

  • The Web UI shows every run, a live terminal for the selected agent, approvals, presets and history, in a light or dark theme and ten languages.

  • A CLI with the MCP attached can ask for approval before a risky step, and you approve or deny it from the CLI or the Web UI.

Develop, review, fix

  • reevesagents flow start attaches a flow to an existing run: a developer agent, a reviewer agent and the repository. Each flow tick moves it one step: the developer gets the next spec, the reviewer checks the commit, and the item is closed or sent back for a fix within the round budgets.

  • Specs wait in a queue (flow add-spec) or come from a plan imported from a document in the repository (flow plan). status, notes, logs and incidents show where each flow stands, what the reviewer found, the journal, and questions or failures that need attention. The TUI, the Web UI and the MCP tools show the same flows, and the MCP server runs a step only when asked.

Agent Control (MCP)

  • reevesagents attach connects every installed host CLI, or one by its id: cc, codex, kimi, qwen, hermes or opencode. hosts shows the state and detach <id> removes it.

  • The host gets tools to spawn agents, read their output, send text and keys, interrupt, stop, and handle approvals and presets. Workers get none of them unless you attach their CLI too.

  • A Claude Code plugin ships in this repository, and the server is listed in the official MCP Registry as io.github.mertkayacs/reevesagents.

Local by design

  • State is plain JSON under ~/.reeves, and REEVES_REGISTRY and REEVES_CONFIG move it elsewhere.

  • Provider credentials are never stored and model traffic goes straight from each CLI to its provider. Known token patterns are redacted before state is written.

  • The Web UI binds to 127.0.0.1, checks the Host header on every request and checks the Origin header on every request that changes state.

Helper tools

  • reevesagents tools lists Perplexity CLI, Google Colab CLI, Hugging Face CLI and Ollama with a purpose, an example command and their install and sign-in state. Agents call them from their own shell.

Quick start

Install with pnpm, check the machine, then open the terminal UI:

pnpm add -g reevesagents
reevesagents doctor
reevesagents

npm, Homebrew and a one-off check work too:

npm install -g reevesagents
brew install mertkayacs/reevesagents/reevesagents
npx reevesagents doctor

Start a team from the command line:

reevesagents spawn claude-code:lead codex:tests hermes:research \
  --name "release check" \
  --prompt "Review the release path, test coverage, and docs."

Open the Web UI, which listens on 127.0.0.1 only:

reevesagents web

Let Claude Code drive the others, then restart it so it loads the MCP tools:

reevesagents attach cc
reevesagents hosts

In Claude Code you can install the connection and its operating instructions as a plugin instead:

claude plugin marketplace add mertkayacs/reevesagents
claude plugin install reevesagents@reevesagents

Requirements

  • Node.js 22.12 or newer.

  • tmux 3.0 or newer on macOS, Linux and WSL. Native Windows 10 1809 or later uses the built-in winmux multiplexer instead.

  • At least one supported coding CLI, installed and signed in. reevesagents doctor checks Node, tmux and which CLIs it can launch.

  • The Web UI uses the optional ws and @lydell/node-pty modules, which npm installs by default.

Supported CLIs

Use the id in a spawn spec such as codex:tests. Aliases like claude work there too, while attach and detach take the id.

CLI

id

MCP host

Claude Code

cc

yes

Codex CLI

codex

yes

OpenCode CLI

opencode

yes

Hermes

hermes

yes

Kimi Code

kimi

yes

DeepSeek CLI

deepseek

Pi

pi

Qwen Code

qwen

yes

Aider

aider

Gemini CLI

gemini

Cline CLI

cline

Crush

crush

Goose

goose

Kilo Code

kilo

Letta Code

letta

Mistral Vibe

vibe

Grok Build

grok

Kiro CLI

kiro

Factory Droid

droid

Qoder CLI

qoder

Warp

warp

Junie CLI

junie

GitHub Copilot CLI

copilot

Cursor CLI

cursor

Rovo Dev CLI

rovo

The binary on your PATH matches the id, except claude (cc), kiro-cli (kiro), qodercli (qoder), cursor-agent (cursor) and acli (rovo). reevesagents providers shows which of them it finds, with their aliases and suggested models.

FAQ

No. Each CLI keeps its own login and keys, the same as when you run it by hand. ReevesAgents keeps run records under ~/.reeves and redacts text that looks like a token before writing it.

Each coding CLI sends your code to its own model provider, as it always does, and ReevesAgents adds nothing in between. The Web UI answers only on 127.0.0.1. To use it from another machine, forward the port over SSH with ssh -L 8080:127.0.0.1:8080 user@host.

No. Agents can work side by side with nobody in charge. Agent Control is an MCP server you attach to one CLI with reevesagents attach cc (or another id), and reevesagents detach cc removes it.

Yes. Codex sandboxes MCP calls by default, which blocks tmux launches, so run it with full access when it hosts the MCP, for example codex --sandbox danger-full-access, or use a Codex profile that sets sandbox_mode = "danger-full-access".

The agents keep working. Each one lives in its own tmux window, or a winmux window on native Windows, and reevesagents open <id> takes you back to it.

Yes. Since 2.1.0 it runs natively on Windows 10 1809 and later through the built-in winmux multiplexer, and it also runs in WSL.

Yes. spawn, runs, agents, providers and doctor accept --json, and REEVES_REGISTRY and REEVES_CONFIG keep a scripted run away from your own ~/.reeves.

ReevesAgents is free under Apache-2.0. Your providers bill you as usual, and ReevesAgents adds nothing on top.

Documentation

Build from source

git clone https://github.com/mertkayacs/reevesagents.git
cd reevesagents
pnpm install
pnpm build
pnpm link --global
reevesagents doctor

Development

Contributors use pnpm, TypeScript, tsup, Vitest and ESLint from the repository. Run the focused checks before a pull request:

pnpm typecheck
pnpm lint
pnpm test
pnpm verify

pnpm verify runs the type check, lint, unit tests, the build, a CLI smoke test against fake provider binaries and a check of the npm package contents. Testing and releasing have the details.

Contributing and security

Issues and pull requests are welcome. See CONTRIBUTING.md. Report vulnerabilities privately as described in SECURITY.md.

If ReevesAgents is useful to you, a ⭐ helps other people find it.

License

Apache-2.0. ReevesAgents is built by Mert Kaya at Eschatia Labs, an independent, non-profit open-source AI research lab. Photos behind the screenshots are credited in docs/assets/readme/CREDITS.md.

views

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers