# OpenBot by Nightly Labs: A Desktop Workspace for Codex, Claude Code and Grok Agents > Source: > Published: 2026-10-26 00:00:00+00:00 **OpenBot**, a repository from nightly-labs, picked up about 312 GitHub stars in a single day this week, and it is easy to see why. It packages the coding-agent CLIs people already use, namely Codex, Claude Code and Grok, into one desktop app where every agent is a persistent teammate with its own workspace, task queue and chat. The [nightly-labs/openbot README](https://github.com/nightly-labs/openbot?ref=explainx) is also unusually blunt about the price of that convenience: agents run with full access and no per-action approval. This post explains what it does, how to try it, what the license allows, and the risks you should weigh first. It is a different project from the [CopilotKit OpenBot we covered in August](https://explainx.ai/blog/copilotkit-openbot-open-source-grok-bot-august-2026), despite the shared name. ## TL;DR | Question | Answer | |---|---| | What is it? | Local-first desktop workspace for persistent AI agents | | Which agents? | Codex App Server and Claude Code; Grok CLI, OpenCode, Gemini, Cursor CLI and Cline CLI through ACP | | Platforms | macOS 13+ (Apple silicon and Intel), Windows 10+ x64, Linux AppImage (x64 and arm64), plus Docker and Nix | | Status | Development preview per the README | | Permissions | Agents run with danger-full-access and approvalPolicy never after first-launch consent | | License | PolyForm Noncommercial 1.0.0, described as not free software | | Do I need API keys? | It reuses your existing local CLI login or provider key; it says it does not copy credentials | | Stars | Roughly 1.2k on the repo page, 312 gained in a day on trending | ## What OpenBot actually does Most multi-agent setups today mean several terminal windows. OpenBot replaces that with one app. According to the README, you get: - **Persistent agents.** Each is backed by an independent session of Codex, Claude, Grok, OpenCode, Gemini, Cursor or Cline and its own local workspace, with editable instructions and an avatar. - **Queues.** Per-agent FIFO message queues with pause, resume, cancel and crash-safe persistence, so you can line up work. - **Agent-to-agent messaging.** Replies, reactions, images and managed file transfers between agents. - **Shared channels.** Desktop channel chats with one task owner, explicit delegation, shared history, and Stop, Resume, Reassign, Archive and Restore controls. - **An embedded browser** that agents can open, inspect and control, plus optional Computer Use on macOS, Windows and Linux through a bundled driver. - **Context monitoring.** Per-agent tracking with automatic compaction before long threads exhaust the model window. - **Custom endpoints.** OpenAI-compatible endpoints and custom ACP agents, with detection of local model servers such as Ollama and LM Studio. - **Local skills.** An agent can create a skill from a workflow, saved to a shared local library that other agents can install. If you have tried parallel agent sessions in [Claude Code desktop](https://explainx.ai/blog/claude-code-desktop-parallel-sessions-sidebar-worktrees-nested-sessions-guide-2026), the idea will feel familiar; the difference is that OpenBot mixes vendors in the same workspace. That puts it in the same family as [AgentPlane, a local control plane for Claude Code, Codex and Cursor](https://explainx.ai/blog/agentplane-local-control-plane-claude-code-codex-cursor-side-by-side-2026) and the [Cindy harness-switching agent](https://explainx.ai/blog/cindy-open-source-agent-claude-code-codex-pi-harness-switching-apache-github-trending-risks-2026). ## How to try it The README lists three routes. **Desktop.** Download the dmg, exe or AppImage from the repository's releases page. On Linux, mark the AppImage executable. On Ubuntu 23.10 or newer and Debian 13, an AppArmor profile (shipped at `build/linux/openbot.apparmor`) must be installed or the app exits at launch, and the README says not to launch with `--no-sandbox`. Windows releases are signed by a Polish company, Synthetify Labs; older releases are not signed, and SmartScreen can still warn, so check the checksum or the GitHub build attestation. **Server.** For an always-on host on Ubuntu 24.04 or 26.04, the README shows a `curl | sudo bash` installer followed by `sudo openbot login`, then use it from the desktop app, an iPhone app or the web client. Read the install script before piping it to root. **Docker.** A container for linux/amd64 and linux/arm64 with a seccomp profile that lets the Electron sandbox stay on, plus `no-new-privileges`. **Agent CLIs.** OpenBot can download and pin supported provider runtimes, or use ones you installed. Gemini uses a Google AI Pro or Ultra plan and downloads Google's Antigravity ACP server separately because of license terms. Cline's free models may use your prompts for training, per the README. OpenCode's free models need no sign-in. For developers the project needs Bun 1.4.2 and Node.js 24. Its seeded dev profile uses a gpt-6-luna model at low effort or an OpenCode-hosted Muse Spark model, which tells you which model families the authors test against. ## The danger-full-access warning This is the part to read twice. The README states: OpenBot is a development preview. Agents currently run with danger-full-access and approvalPolicy: never. It continues that agents "can read and modify files, run commands, use the network, and control the embedded browser without per-action confirmations after the explicit first-launch consent." The advice is to run only agents and tasks you trust, keep backups and review the Security section. What that means in practice: - **A prompt injection becomes a shell.** If an agent browses a page or reads an issue containing hostile instructions, it can act on your machine with no prompt. Our write-up of[an agent-security red-teaming study of auto-mode monitors](https://explainx.ai/blog/anthropic-red-teaming-auto-mode-codex-guardian-monitors-fail-55-percent-attacks-2026) shows why classifier-style guardrails are not a safe default. - **Agents talking to agents widen the blast radius.** A compromised or confused agent can message another agent that has different tools or credentials. - **Shared channels add shared state.** Delegation across agents is convenient, and it also makes it harder to say who did what. - **Computer Use and the embedded browser** can click through logged-in sessions. Sensible mitigations: run it in a VM or the Docker image, give each agent a throwaway workspace, avoid mounting your home directory, use scoped API keys, and keep production credentials out of the environment. For a runtime view of what agents attempt, teams can look at [AgentBeam](https://agentbeam.com/?ref=explainx), the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions. ## License: read it before using it at work The README badge shows PolyForm Noncommercial 1.0.0, and the Nix section says plainly that OpenBot is not free software and must be allowed as an unfree package. The [PolyForm Noncommercial license](https://polyformproject.org/licenses/noncommercial/1.0.0/?ref=explainx) permits noncommercial uses, such as personal study, and does not give commercial rights. A company that wants to run it for paid work should get terms from the copyright holder. We did not independently audit older releases or verify any earlier licensing; check the LICENSE and CLA files in the repository before you rely on any version. That puts OpenBot in a different category from open-source harnesses with permissive licenses. If you need a permissively licensed alternative, see our guides to [OpenCode 2.0](https://explainx.ai/blog/opencode-2-0-code-mode-background-agents-shared-server-sdk-upgrade-guide-2026) and the [Archestra self-hosted agent runtime](https://explainx.ai/blog/archestra-1-4-arpa-openappa-guardrails-remote-agent-runtime-rbac-claude-code-codex-self-hosted-release-notes-2026). ## Terms of service and data questions Several questions follow from using vendor CLIs inside a third-party app. - **Provider terms.** The README says each provider's terms apply to your sign-in and points to a provider-terms document for Claude subscriptions. If you run agents through a Claude subscription inside a wrapper, check Anthropic's current terms on automated use. Using official API keys is the safer path. - **Credentials.** OpenBot says it uses the existing local CLI login and does not copy provider credentials, and that Grok keys and per-session bearer tokens are never persisted or logged. That is a claim to verify by reading the code or by network monitoring, not a guarantee. - **Telemetry.** The repository ships ANALYTICS.md and PRIVACY.md. Read both. Optional accounts use one-time email codes with an API on Cloudflare Workers and D1. - **Network traffic.** "Local-first" means your agent state is local. Prompts and code still go to OpenAI, Anthropic, xAI, Google, Cursor or Cline depending on the agent. For an example of the grey-area side of this ecosystem, see our piece on the [free-claude-code proxy and its terms and data risks](https://explainx.ai/blog/free-claude-code-proxy-61-providers-1-3b-free-tokens-terms-data-risks-2026). ## How it compares | Tool | Shape | Vendors | License | |---|---|---|---| | OpenBot (nightly-labs) | Desktop app plus optional server | Codex, Claude Code, Grok, OpenCode, Gemini, Cursor, Cline | PolyForm Noncommercial | | AgentPlane | Local control plane | Claude Code, Codex, Cursor | see our write-up | | Cindy | Open-source agent switching harnesses | Claude Code, Codex, Pi | Apache per our write-up | | Claude Code desktop | Vendor app with parallel sessions | Claude only | Vendor terms | We have not benchmarked these against each other, and no performance claims for OpenBot appear in the README. The case for it is workflow, not speed. ## What people are asking **Do I need an account?** No, accounts are optional and use email codes. You do need logins or keys for the agents you enable. **Will it work on Wayland?** The README says remote desktop on Linux needs an X11 session and does not work under Wayland, and the arm64 AppImage omits it. **Can I import agents from elsewhere?** The README describes importing agents from Grok Bot via an export agent that produces a zip of instructions, skills, routines and memories; chat history is not copied. **Is it ready for a team?** It is labelled a development preview, with 139 open issues and 37 open pull requests on the repo page when we checked. Treat it as an experiment. ## What this means if you build with agents OpenBot is a signal that the harness layer is moving from terminals to workspaces, where vendors are interchangeable and the interesting objects are queues, channels and shared skills. That is useful for individuals who juggle several agents. For teams, the license, the permission model and the provider terms are the real evaluation criteria, ahead of features. A cautious first trial: install the Docker image on a spare machine, connect one agent with a scoped API key, give it a disposable repository, and watch what it does with the embedded browser before you add more agents. *Details reflect the repository README and page as of October 26, 2026; star counts and features change quickly.*