Cowork – Self-hosted task queue and dashboard for Claude, Codex, and AI

An open-source developer named slashman413 published Cowork, a self-hosted control plane and Model Context Protocol server that coordinates coding agents across disparate command-line tools without an external database. According to github.com's report, the project gives local developer tools a unified task inbox, a hardware telemetry dashboard, and automated fallback chains that redirect work whenever an individual provider limits execution quota.

The software addresses engineering teams and independent developers who juggle terminal-based AI agents across distinct models and runtimes. By consolidating session states into simple disk files, Cowork allows engineers running native command-line sessions in Claude Code, Gemini CLI, Hermes Agent, and Antigravity to retain working task context when switching between cloud API endpoints and local model hosts.

01Fallback routing across local and remote engines

Cowork structures automated execution chains that route developer prompts sequentially from Claude Code to Hermes, then Gemini, and finally to local Ollama runtimes. When a primary model API returns a rate limit error, the control plane shifts the task context to the designated backup without clearing active git patches or developer instructions. This mechanism prevents developers from stalling midway through development sprints when cloud quotas expire.

External compute nodes join the local hub through a one-line Node.js command, registering hardware directly into the central task dispatcher. Developers set the target hub address and host identifier to connect remote Mac Studios, secondary laptops, or cloud virtual machines over network endpoints. This arrangement allows distributed machines to contribute local graphics processing units or specialized models directly to the central inbox.

02Single pane dashboard and zero database storage

The control plane pairs its Model Context Protocol server with a dark-themed operational web dashboard that presents host hardware telemetry. The overview interface monitors processor utilization, graphics chip loads, working memory usage, and component temperatures alongside live agent statuses. It also surfaces a seven-day quality panel tracking task completion rates, retry counts, input pauses, and operational timeouts.

Unlike multi-agent frameworks that require relational databases, Cowork eliminates database management by storing all system records in plain JSON files and Markdown documents. Task queues, status logs, operational decisions, and generated artifacts reside within dedicated local file directories. This zero-database architecture avoids schema migrations, background daemon dependencies, and complex data container setups across local environments.

03Integration packs and platform agent formatting

Cowork organizes specialized agent behaviors into two hundred and eighty-five built-in specialist personas distributed across nineteen operational departments. These personas cover domain specializations including frontend engineering, site reliability operations, security review, penetration testing, and data infrastructure. The system automatically fetches integration packs from designated repositories on first launch to populate the dispatcher's routing roster.

The project translates agent definitions into native file formats across varied programming assistants and editor extensions. Personas map to Markdown documents with YAML frontmatter for Claude Code and Copilot, formatted skill files for Hermes and Antigravity, TOML definitions for Codex, and MDC rule files for Cursor. This format translation allows identical role behaviors to deploy across seven separate developer environments without manual schema conversion.

04Node environment setup and containerized deployment

Deploying the Cowork hub requires Node.js version 20 or higher, with automated testing conducted on version 22, alongside npm version 10. The recommended installation uses two deployment scripts that compile the server application, establish a user systemd service, and inject the coordination skill into local agent directories. The configuration installer non-destructively merges connection parameters into local client files such as the primary Claude configuration.

Teams opting for containerized infrastructure can run prebuilt images distributed through the GitHub container registry under version 1.0.1. The container image runs the task inbox, dashboard interface, and streamable server on network port 6868 while mapping persistent state to a dedicated data volume. The release workflow produces multi-architecture container images supporting both amd64 and arm64 systems with digest-pinned base images.

05Network security settings and client access controls

The hub manages network communication through local configuration files stored at user home paths, preventing private network settings from committing to git repositories. Server parameters define network interface bindings, listening ports, and authentication credentials. When binding beyond loopback addresses without a preconfigured key, the server generates a random authentication token on first initialization and prints it once to setup logs.

To protect host machines, Cowork incorporates an explicit client guard that screens incoming traffic against network addresses and client identifiers. The server returns forbidden status codes to matched items found in user agent headers, connection payloads, or tool parameters regardless of key validity. Flagged agents are automatically removed from active routing chains, ensuring misbehaving or untrusted workers cannot pull tasks from the queue.