Cowork: The Open-Source Task Queue That Lets Claude Code, Codex and Gemini CLI Share One Inbox

If you code with AI agents, you know the moment: two hours into a sprint, Claude Code stops cold with a rate-limit error, and your whole workflow stalls. You switch to Gemini CLI or a local model, but the task state, the half-finished patch, and the context you built up do not come with you. Cowork, an open-source project that surfaced in AI news on November 14, 2026, exists to fix exactly that. It is a self-hosted task queue and dashboard that puts Claude Code, Codex, Gemini CLI, Hermes, and other coding agents into one shared inbox — so when one agent hits its limit, the work simply moves to the next one.

First, a name check, because the branding is confusing. Cowork is not Anthropic’s Claude Cowork, the cloud-hosted agentic work mode that runs inside your Claude account. This Cowork is an independent open-source project by developer slashman413, published under the MIT license on GitHub. The similarity ends at the name — and, fittingly, at the idea that an agent should keep working while you do something else.

What Cowork Actually Does

At its core, Cowork is a control plane for coding agents. Instead of each agent living in its own terminal window with its own to-do list, every task lands in a single JSON inbox on a self-hosted hub. A dispatcher routes each task in two stages — first to a division (frontend, SRE, security, data engineering, and 15 more), then to a specialist persona — and executes it along an agent fallback chain. The project ships with more than 285 built-in specialist personas across 19 departments, so a security review and a database migration can be routed to different agent identities with different instructions.

The hub itself is an MCP server plus a dark-themed web dashboard, served on port 6868. The dashboard shows live host metrics such as CPU, GPU, memory, and temperature, active workers, open tasks, a seven-day quality panel with pass rates and retries, and a real-time activity feed. There are also views for scheduled tasks and recurring routines, long-running goals with phases and budgets, saved recipes you can re-run with one click, and an integrations page that installs skills, MCP servers, and persona packs from audited Git repositories.

The Killer Feature: Fallback Chains, Not Retries

The idea that makes Cowork worth a look is the fallback chain. A typical chain reads Claude Code, then Hermes, then Gemini, then Ollama. When the first agent in the chain fails, hits a quota, or a remote machine leaves a task unclaimed past its grace window, the task hands over to the next agent in line — and every attempt is recorded on the task card. Nothing is silently dropped, and nothing is retried blindly against the same exhausted quota.

Three robot assistants passing a glowing task token along a chain, illustrating agent fallback

This matters because the failure mode of agent coding is rarely a bad model — it is a good model that is temporarily unavailable. Anyone who has watched a long agent run die at 2 a.m. because of a usage cap understands the appeal of a queue that simply tries the next brain. It is the same instinct behind LangChain’s $300M bet on turning LangSmith into the agent operating system: as agents move from demos into production, the boring infrastructure around them — queues, retries, observability — becomes the product.

One Inbox for Claude Code, Codex, Gemini CLI, and More

Cowork connects to agents through the Model Context Protocol, so any MCP-compatible client can create, claim, and complete tasks. The README lists native support for Claude Code, Codex, Google’s Gemini CLI and Antigravity, the Hermes agent, GitHub Copilot, and Cursor, plus Ollama for local models. A task can be dispatched from any of them — or from the dashboard UI — and a dispatcher routes it to a named “brain,” meaning a specific combination of model, CLI, and machine.

Remote machines join the fleet with a one-line Node.js client, and claims are atomic, so two workers can never double-run the same task. That makes Cowork usable as a small home cluster: a laptop, a Mac Studio, and a cloud box can all pull from the same inbox. It is a deliberately different shape from Snowflake’s CoCo and CoWork bots, which are managed, always-on coworkers inside Snowflake’s platform — Cowork is the do-it-yourself, self-hosted version of the same always-on idea, and despite the similar name it is an unrelated project.

Self-Hosted Means You Own the Whole Stack

Unlike managed agent platforms, Cowork keeps every task, prompt, and response on your own infrastructure. The task store is plain JSON and Markdown files on disk — no Postgres, Redis, or SQLite to install, back up, or migrate. The dashboard runs locally and needs no telemetry or external connections beyond the AI model APIs themselves.

Home server rack and laptop running a dark AI agent dashboard

For teams handling sensitive code, working under compliance rules, or just uncomfortable with a third-party service seeing their prompts, that is the point. It is also the trade-off: you are the operator now. You install it, you update it, you secure the dashboard, and you pay the model bills — Cowork itself charges nothing for the core, though a separate Cowork Pro package ($59) adds advanced workflows, priority support, and lifetime updates.

Getting Started in About 30 Seconds

The project advertises a three-line quickstart, and the README backs it up. The prerequisites are modest: Node.js 20 or newer and npm 10. From there it is:

git clone https://github.com/slashman413/cowork && cd cowork
deploy/install-server.sh    # build and launch the MCP server + dashboard
deploy/install-skill.sh     # wire the Cowork connection into your agent clients

Then open http://localhost:6868, create a task, and watch the dispatcher route it. Prefer containers? The project publishes an image you can run with a single Docker command: docker run -d -p 6868:6868 -v cowork-data:/data ghcr.io/slashman413/cowork:1.0.1. On first run, the server provisions default integrations in the background, including persona packs and skill collections from public repositories, which you can inspect or override in the configuration.

Who Cowork Is For — and Who Should Skip It

Cowork makes the most sense for developers and small teams who already run two or more coding agents and keep tripping over rate limits, and for homelab users who enjoy owning their tooling. If you live inside a single agent and rarely hit a quota, a shared inbox buys you little. If your team needs audit trails, SSO, or a vendor to call at 3 a.m., a managed platform is still the safer choice.

There is also an honest maturity caveat. The GitHub repository has been public since July 2026 and carries a few hundred commits, but it is still a small project — four stars and a single fork at the time of writing — with no formal releases, only tags. That is normal for an early open-source tool, but it means you should read the code before you trust it with production work, and expect the rough edges that come with a one-maintainer project. The upside of open source is that you can do exactly that: every file, every routing rule, and every skill is visible in the repo.

The bigger picture is that agent infrastructure is having its moment. OpenAI is teaching ChatGPT’s agents to pay for purchases on your behalf, its “dots” now work across multiple computers, and Brave lets agents pay per search with crypto. Agents are becoming economic actors, not just chatbots — and economic actors need queues, fallbacks, and dashboards. Cowork is a small, free, self-hosted answer to that need. If you run agents in production, it is worth an hour of your weekend.

Frequently Asked Questions

Is Cowork the same as Anthropic’s Claude Cowork?

No. Anthropic’s Claude Cowork is a cloud-hosted agentic work mode inside Claude, available on paid Claude plans. Cowork is an independent open-source project by slashman413 that you host yourself. They share a name and a similar goal — agents that keep working while you are away — but they are unrelated products.

Which AI agents does Cowork support?

Claude Code, Codex, Gemini CLI and Antigravity, Hermes, GitHub Copilot, Cursor, Ollama, and any MCP-compatible client. Tasks are routed to a named “brain” — a model, CLI, and machine combination — and fall back to the next agent in the chain on failure or rate limits.

Does Cowork need a database?

No. The task store is plain JSON and Markdown files on disk. There is no Postgres, Redis, or SQLite to maintain, which keeps deployment to a single Node.js service or one Docker container.

How much does Cowork cost?

The core server, dashboard, and remote-agent client are free and open source under the MIT license. A separate Cowork Pro package costs $59 and adds advanced workflows, priority support, and lifetime updates. You still pay your own AI model or subscription bills.

How do I try Cowork?

Clone the repository from GitHub, run the two install scripts, and open http://localhost:6868 — or run the published Docker image with a single command. You will need Node.js 20 or newer.

Sources: Cowork on GitHub; Emergent — Cowork: Self-Hosted Task Queue for Claude and AI Agents; AiNews247 — Cowork, self-hosted task queue and dashboard for Claude, Codex, and AI agents; Glama MCP listing for Cowork.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top