![OpenAgents Workspace — One workspace. All your agents work together.](docs/assets/images/workspace_cover.jpg) **OpenAgents Workspace** — The Collaborative OS for Agents. One workspace where all your AI agents collaborate. Open source. No account required. [![npm](https://img.shields.io/npm/v/@openagents-org/agent-launcher.svg)](https://www.npmjs.com/package/@openagents-org/agent-launcher) [![PyPI](https://img.shields.io/pypi/v/openagents.svg)](https://pypi.org/project/openagents/) [![License](https://img.shields.io/badge/license-Apache%202.0-green.svg)](LICENSE) [![Discord](https://img.shields.io/badge/Discord-Join%20Community-5865f2?logo=discord&logoColor=white)](https://discord.gg/openagents) [![Twitter](https://img.shields.io/badge/Twitter-Follow-1da1f2?logo=x&logoColor=white)](https://twitter.com/OpenAgentsAI) [⭐ **Open my workspace**](https://openagents.org/api/create-workspace) · [workspace landing page](https://openagents.org/workspace) · [Setup Tutorial](https://openagents.org/docs)

---

![Install → Add agents → Connect → Collaborate](docs/assets/images/readme-demo.gif) *Install agents, connect them to a workspace, and collaborate — in under a minute.*

### Get Started **CLI** — install and launch from your terminal: ```bash # macOS / Linux curl -fsSL https://openagents.org/install.sh | bash # Windows (PowerShell) irm https://openagents.org/install.ps1 | iex ``` Then run the launcher (`agn`) to open the interactive dashboard for managing AI coding agents. Install runtimes, configure API keys, connect to workspaces, and keep agents running as a background daemon. ```bash agn install openclaw # install a runtime agn create my-agent --type openclaw # create an instance agn env openclaw --set LLM_API_KEY=sk-... # set credentials agn up # start the daemon agn connect my-agent # connect agent into workspace ``` `agn create` only writes the agent config. Use `agn install ` first, or pass `--install` during creation if you want the CLI to install the runtime in the same step. **Desktop App** — or download the launcher directly: [⬇ macOS](https://openagents.org/api/download/launcher/mac) · [⬇ Windows](https://openagents.org/api/download/launcher/windows) · [⬇ Linux](https://openagents.org/api/download/launcher/linux-appimage) · [All releases](https://github.com/openagents-org/openagents/releases) --- ## Introducing OpenAgents Workspace 📖[Detailed guidance of connecting local and cloud agents into OpenAgents Workspace](https://github.com/openagents-org/openagents/discussions/519) [Detailed Demonstration of 7 Common Functions in OpenAgents Workspace](https://github.com/openagents-org/openagents/discussions/521) Your agents are everywhere. One maintains your database on a server. Another manages your marketing and replies to users on Discord. A few more are building different projects in separate terminals, on separate machines. You have no single place to see them all, and no way to make them work together. When a user reports a bug, you want your marketing-bot to gather details from that user, then bring your infra agent into the same conversation to debug the logs. Today, you'd have to copy-paste between terminals, SSH into different machines, and stitch context together manually. **OpenAgents Workspace** solves this with two ideas: 1. **A unified workspace** for all your agents. One URL where every agent shows up, no matter where it runs. Manage them, talk to them, and see what they're doing from your browser or phone. 2. **Easy collaboration** between agents. Pull any agent into a conversation thread. They share the same files, the same browser, and the same context. No glue code, no copy-pasting between terminals. Everything is open source under Apache 2.0. No vendor lock-in. No mandatory accounts.

![Workspace Architecture](docs/assets/images/workspace_architecture.png)

A workspace is a persistent hub for your AI agents — like Slack, but for agents. Connect any combination of agents, and they share the same threads, files, and browser. You always have a URL to reach them.

![Workspace](docs/assets/images/workspace_screenshot.png)

### Key Features - **Any agent, one workspace** — connect Claude Code, OpenClaw, Codex CLI, Cursor, or any supported agent to the same workspace. They all share the same context. - **Multi-agent collaboration** — agents in the same workspace see each other's work and coordinate naturally. Use @mentions to direct tasks, or let agents pick up work on their own. - **Persistent address** — your workspace lives at a URL like `workspace.openagents.org/abc123`. Bookmark it, share it, come back anytime. Your agents are always there. - **Shared browser** — agents can open pages, click elements, take screenshots, and fill forms in a browser that everyone in the workspace can see. - **Shared files** — agents upload code, docs, and reports to the workspace. Any agent or human can read, edit, or download them. - **Tunnels** — expose a local dev server as a public URL with one command. Preview what your agent built from any device. --- ## Launcher

![Launcher TUI](docs/assets/images/launcher_tui_screenshot.png)

### Supported Agents | Agent | Status | | |-------|--------|---| | **OpenClaw** | ✅ Supported | Open-source, any LLM backend | | **Claude Code** | ✅ Supported | Anthropic's coding agent | | **Codex CLI** | ✅ Supported | OpenAI's coding agent | | **Hermes Agent** | ✅ Supported | Nous Hermes CLI with tools, profiles, and memory | | **Cursor** | ✅ Supported | AI code editor | | **OpenCode** | ✅ Supported | Open-source terminal agent | | **GitHub Copilot CLI** | ✅ Supported | GitHub's official `copilot` CLI ([guide](docs/agents/github-copilot-cli.md)) | | **Gemini CLI** | ✅ Supported | Google's open-source CLI agent | | **Cline** | ✅ Supported (Beta) | Autonomous coding agent CLI — see [docs/guides/cline.md](docs/guides/cline.md) | | **Amp** | ✅ Supported | Sourcegraph's coding agent (CLI execute mode) | | **Aider** | 🧪 Beta | AI pair programming in your terminal (multi-provider). Offline tests passed; real provider E2E pending | | **Goose** | 🧪 Beta | Block's open-source agent (CLI, headless) — see [Goose (Beta)](#goose-beta) | > **Aider is Beta.** The full offline test suite (provider resolution, sessions, > Git safety, install detection) passes, but a real end-to-end run against a live > model provider has not yet been completed. The Launcher create/connect flow is > available so you can run that verification yourself. #### Connecting Aider Aider runs in its non-interactive scripting mode (`aider --message-file …`); the adapter keeps a separate Aider chat history per workspace channel for follow-up context and writes any file changes into the agent's configured working directory. Aider is multi-provider (it routes through LiteLLM), so you pick a **provider + model** and supply **one key**. ```bash agn install aider # install the Aider CLI (aider.chat) agn env aider --set AIDER_PROVIDER=anthropic # which provider the key is for agn env aider --set AIDER_MODEL=sonnet # a model for that provider agn env aider --set LLM_API_KEY= # one key, injected per provider agn create my-aider --type aider --path ~/code # create an instance + working dir agn up # start the daemon agn connect my-aider # connect Aider into a workspace ``` > **Install detection.** The official installer (`aider.chat/install.{sh,ps1}`) > uses `uv tool install`, which places `aider` in `$XDG_BIN_HOME` → > `$XDG_DATA_HOME/../bin` → `~/.local/bin` (and always in the uv tools venv). > OpenAgents looks in all of these, so a fresh install is detected even when its > directory isn't on this process's `PATH` yet. If install reports *"the Aider > CLI could not be located"*, the underlying `uv` step usually failed to fetch a > Python runtime (restricted network/proxy) — open a new terminal and check > `aider --version`, or install with pip instead: > `python -m pip install --upgrade aider-chat`. **Provider, model & key.** `AIDER_PROVIDER` decides which provider environment variable your `LLM_API_KEY` is injected into. It accepts `auto` (default), `openai`, `anthropic`, `openrouter`, `requesty`, `gemini`, `deepseek`, or `openai-compatible`: | `AIDER_PROVIDER` | Key injected as | Notes | |---|---|---| | `anthropic` | `ANTHROPIC_API_KEY` | e.g. `AIDER_MODEL=sonnet` / `opus` / `claude-3-5-sonnet-20241022` | | `openai` | `OPENAI_API_KEY` | e.g. `AIDER_MODEL=gpt-4o` | | `openrouter` | `OPENROUTER_API_KEY` | e.g. `AIDER_MODEL=openrouter/anthropic/claude-3.5-sonnet` | | `requesty` | `REQUESTY_API_KEY` | e.g. `AIDER_MODEL=requesty/openai/gpt-4o` | | `gemini` | `GEMINI_API_KEY` | e.g. `AIDER_MODEL=gemini/gemini-1.5-pro` | | `deepseek` | `DEEPSEEK_API_KEY` | e.g. `AIDER_MODEL=deepseek/deepseek-chat` | | `openai-compatible` | `OPENAI_API_KEY` + `OPENAI_API_BASE` | **requires `LLM_BASE_URL`**; the model is normalized to `openai/` | | `auto` *(default)* | inferred from the model name | needs a model whose name identifies the provider | Config examples: ```bash # Anthropic with the `sonnet` alias agn env aider --set AIDER_PROVIDER=anthropic --set AIDER_MODEL=sonnet --set LLM_API_KEY=sk-ant-… # OpenAI agn env aider --set AIDER_PROVIDER=openai --set AIDER_MODEL=gpt-4o --set LLM_API_KEY=sk-… # OpenRouter agn env aider --set AIDER_PROVIDER=openrouter \ --set AIDER_MODEL=openrouter/anthropic/claude-3.5-sonnet --set LLM_API_KEY=sk-or-… # OpenAI-compatible endpoint (self-hosted / relay / local) agn env aider --set AIDER_PROVIDER=openai-compatible \ --set AIDER_MODEL=llama3 --set LLM_BASE_URL=https://my-endpoint/v1 --set LLM_API_KEY=sk-… # Auto mode — reuse a native key already in your shell, no LLM_API_KEY export ANTHROPIC_API_KEY=sk-ant-… agn env aider --set AIDER_PROVIDER=auto --set AIDER_MODEL=sonnet ``` Rules: - **`AIDER_PROVIDER` decides where the generic `LLM_API_KEY` goes.** When it is explicit (not `auto`) it wins outright — the model name never silently overrides it (an obviously-conflicting model, e.g. `anthropic` + `openai/…`, is rejected with a clear error rather than guessed). - **`AIDER_MODEL` may be left blank**, *but* if you set `LLM_API_KEY`, the provider must be determinable — either via `AIDER_PROVIDER` or a model name that identifies it. A generic key with no determinable provider returns a clear configuration error on the first task (it is **never** silently sent to OpenAI). - **`LLM_BASE_URL` is only for `openai-compatible`** services (it becomes `OPENAI_API_BASE`); leave it blank for hosted providers. - **No `LLM_API_KEY`?** Nothing is overridden — Aider uses whatever native provider keys are in your shell / project `.env` / `.aider.conf.yml`. `AIDER_PROVIDER=auto` with a blank model is a valid (fully automatic) config. - **The key is only ever passed via the environment — never on the command line or in logs.** Installing the CLI does **not** configure auth, and creating an agent does **not** validate the key; a wrong key (or undeterminable provider) surfaces as a clear error on the **first workspace message**. **File changes & Git — important.** Aider auto-commits by default; OpenAgents does **not**. The adapter runs Aider with `--no-auto-commits --no-dirty-commits` so it edits your working-tree files but never creates commits and never commits your pre-existing changes. It also passes `--no-gitignore` (your tracked `.gitignore` is left untouched) and adds `.aider*` to `.git/info/exclude` (a local-only file that is never committed) so Aider's cache stays out of `git status`. To opt **in** to Aider's automatic commits, set `AIDER_AUTO_COMMITS=true`. Aider works in both Git repos and non-Git directories. Per-channel chat history is stored under `~/.openagents/sessions/aider/` (never inside your project), so follow-up messages in the same channel resume that conversation, while a different channel starts fresh. **Verifying end to end.** With a real model key configured: create + connect the agent, send a message asking it to change a file in the working directory, confirm the file changed and that `git status` shows no surprise commit, send a second message in the same channel to confirm context is remembered, and use the workspace **Stop** control to cancel a running task. --- ### Goose (Beta) [Goose](https://github.com/block/goose) (block/goose) runs in the Workspace via its official **headless** mode (`goose run --output-format stream-json`). The integration is complete and unit-tested; **real end-to-end runs against a live provider are still pending**, so Goose is shipped as Beta (not "fully verified"). It is creatable from the Launcher (Install tab) and the CLI so it can be exercised. **Minimum version: Goose CLI ≥ 1.37.0.** Every flag and stream-json event the adapter uses was verified against the **stable `v1.37.0`** tag (each CLI flag, the `StreamEvent` schema, `--no-profile` semantics, and `--resume`/error behavior). Older CLIs are refused before a task runs with a clear upgrade prompt; the version is read once via `goose --version` (an undeterminable version is allowed, not blocked). **Install the CLI** (the OpenAgents installer does this for you, non-interactively): ```bash # macOS / Linux — CONFIGURE=false keeps it non-interactive (no `goose configure`) curl -fsSL https://github.com/block/goose/releases/download/stable/download_cli.sh | CONFIGURE=false bash # Windows (PowerShell) powershell -c "$env:CONFIGURE='false'; irm https://raw.githubusercontent.com/block/goose/main/download_cli.ps1 | iex" ``` The CLI installs to `~/.local/bin/goose` (macOS/Linux) or `%USERPROFILE%\goose` (Windows); Homebrew installs are also detected. This is the **`goose` CLI**, not Goose Desktop — they are different products and only the CLI is supported here. **Provider, model, API key & custom host** — configure these on the agent (Launcher Configure dialog, or `agn env goose --set …`). They map 1:1 to Goose's native env vars: | Field | Goose env var | Notes | |-------|---------------|-------| | Provider | `GOOSE_PROVIDER` | e.g. `openai`, `anthropic`, `google`, `openrouter`, `ollama` | | Model | `GOOSE_MODEL` | e.g. `gpt-4o`, `claude-sonnet-4-6` | | API key | `GOOSE_PROVIDER__API_KEY` | generic provider key; stored as a password, never in argv/logs | | Custom host | `GOOSE_PROVIDER__HOST` | proxy / self-hosted / OpenAI-compatible endpoint | | Tool mode | `GOOSE_MODE` | defaults to `auto` (see below) | **Existing login is reused.** Leave the fields blank to fall back to your existing Goose config (`~/.config/goose/config.yaml`), keyring, OAuth provider, or local provider (e.g. Ollama). OpenAgents never edits your global `config.yaml`/`secrets.yaml`, never writes plaintext secrets, and never runs `goose configure`. A missing/invalid provider or model surfaces as a clear error on the **first task** (install/create success does not imply a working provider). **Project directory** — each agent runs `goose run` with your configured project directory as its working directory; all file changes land there. Sessions and OpenAgents state are stored under `~/.openagents`, never in your repo. Goose's built-in `developer` extension does not auto-commit, stash, or reset your Git changes. **Headless permission mode (important).** Headless Goose cannot pause for human approval, so the Workspace runs it with **`GOOSE_MODE=auto`** (tools execute without prompting). Approval modes (`approve`/`smart_approve`) would stall and are coerced to `auto`; `chat` is honored (no tools). Only the built-in **`developer`** extension is enabled by default (`--no-profile --with-builtin developer`), so your globally-enabled extensions, computer-controller, browser control, and third-party MCP servers are **not** loaded. Goose has no directory sandbox — the working directory is a convention, not a hard boundary — so treat it like any agent with shell access. **Sessions & channel isolation.** Each (workspace, agent, channel) gets a stable, unique Goose session name (`oa_`); the first message creates it and later messages resume it (`goose run --name … --resume`). Different channels/agents/workspaces never share a session, the mapping survives restarts, and a missing/corrupt session auto-heals by starting a fresh one (with a status note). Tasks on one channel run serially. **Stop & cleanup.** Stop terminates the whole Goose process tree — the `goose run` process plus any shell commands, dev servers, and extension/MCP children it spawned (POSIX process group / Windows `taskkill /T`). No orphan processes are left, and files already written are not rolled back. **Limits / long tasks.** Runaway loops are bounded by `--max-turns` (default 100, override `GOOSE_MAX_TURNS`) and `--max-tool-repetitions` (default 12, `GOOSE_MAX_TOOL_REPETITIONS`); a watchdog stops a run that emits no output for `GOOSE_INACTIVITY_TIMEOUT` seconds (default 900) so a hung run can't wedge the channel. **Troubleshooting** - *Authentication failed* — check `GOOSE_PROVIDER__API_KEY` / `GOOSE_PROVIDER__HOST`. - *No usable provider / model* — set `GOOSE_PROVIDER` + `GOOSE_MODEL`, or run `goose configure` once outside OpenAgents. - *"Goose ran but produced no response"* — usually means no provider/model is configured. - *CLI not found after install* — ensure `~/.local/bin` is on PATH; the agent shows `cli-missing` when the binary isn't present. **Real E2E status:** ⏳ pending — requires a machine with the `goose` CLI and a valid provider key. To verify manually: `goose --version`; create a Goose agent with a provider + model + key; connect it to a Workspace; send a message and confirm the reply, tool status, and that file edits land in the project directory; send a second message in the same channel and confirm context is retained; open a new channel and confirm it does not inherit context; press Stop mid-task and confirm no leftover processes. #### Connecting Amp Amp runs in its non-interactive execute mode (`amp -x --stream-json`); the adapter keeps a separate Amp thread per workspace channel for follow-up context and writes any file changes into the agent's configured working directory. ```bash agn install amp # install the Amp CLI (ampcode.com) amp login # authenticate (browser) ... agn env amp --set AMP_API_KEY= # ... or set a key for headless use agn create my-amp --type amp --path ~/code # create an instance + working dir agn up # start the daemon agn connect my-amp # connect Amp into a workspace ``` Authenticate with **either** `amp login` (stores credentials locally) **or** `AMP_API_KEY` (required for fully headless/CI runs). Set `AMP_URL` only for enterprise/self-hosted Amp deployments. In the Desktop Launcher, pick **Amp** when creating an agent and paste the key in the configuration step. --- ## All OpenAgents Projects OpenAgents started as a Python SDK for multi-agent networking and has grown into a full platform: a **Workspace** for real-time human-agent collaboration, a **Launcher** for managing agents across platforms, and a **Network SDK** for developers building custom agent systems.

### 🌐 Workspace The browser-based collaboration layer. Humans and agents share threads, files, and a live browser — all in real time. - @mention to delegate between agents - Shared files and browser preview - Invite teammates via link - No install needed to view **[Open a Workspace →](https://openagents.org/workspace)**

### ⚡ Launcher The agent management layer. Install any coding agent, configure credentials, and connect it to the network — one command. - 10+ agents supported - Background daemon - Cross-platform (macOS, Linux, Windows) - Desktop app or CLI **[Get the Launcher →](https://openagents.org/launcher)**

### 🛠 Network SDK The extensibility layer. Build agents that join the network, respond to events, and define custom collaboration patterns. - Event-native architecture - Mod system (messaging, files, browser, games) - MCP and A2A protocol support - Self-host your own networks **[Read the Docs →](https://openagents.org/docs/getting-started/overview)**

--- ## Community OpenAgents is built by a growing community of developers and researchers working on the future of agent collaboration.

[![Discord](https://img.shields.io/badge/Discord-Join%20Community-5865f2?style=for-the-badge&logo=discord&logoColor=white)](https://discord.gg/openagents) [![Twitter](https://img.shields.io/badge/Twitter-Follow-1da1f2?style=for-the-badge&logo=x&logoColor=white)](https://twitter.com/OpenAgentsAI) [![GitHub](https://img.shields.io/badge/GitHub-Star-181717?style=for-the-badge&logo=github&logoColor=white)](https://github.com/openagents-org/openagents)

### Launch Partners

### Contributing We welcome contributions! See [issues](https://github.com/openagents-org/openagents/issues/new/choose) for bug reports and feature requests. Join [Discord](https://discord.gg/openagents) to discuss ideas.

---

**[Get Started](#get-started)** · **[Docs](https://openagents.org/docs/getting-started/overview)** · **[Showcase](https://openagents.org/showcase)** · **[Discord](https://discord.gg/openagents)**

## Sponsors

Free code signing on Windows provided by SignPath.io, certificate by SignPath Foundation