The value is context, not just speed: work that would fill a single session gets split so each agent carries only its own slice. Plan and execute are separable — you can get the task breakdown, dependencies and estimates back without anything running yet. Note what it needs before you commit: the agents launch the Claude CLI with permissions bypassed, which is a decision to make deliberately.
An orchestration layer over the Claude Code CLI. It breaks a problem into tasks, runs them across parallel agents with dependencies respected, and aggregates the results.
- A problem decomposed into a structured task breakdown before anything runs
- A plan you can read first — tasks, dependencies and time estimates — returned without executing it
- Parallel execution across the mesh with dependency management and result aggregation, bounded by a concurrency limit you set
- End-to-end solving in one call, where the analysis and the execution happen together for a chosen approach and agent count
- Single specialized agents spawned for one task with a named role, returning an agent ID immediately
- Progress, completion status and time estimates for the agents in a session, and results collected per agent when they finish
- Network status: which agents are active, what has completed, and the performance metrics behind it
The Claude Code CLI installed and already authorized: run claude --dangerously-skip-permissions once by hand, log in and accept the terms, or the server's agents cannot start. That flag is how the spawned agents run, so they execute without permission prompts. Node.js v20 or newer. Clone the repository, npm install and npm run build, then point your client at dist/mesh-server.js; the npm package is @thomaswalichiewicz/meshseeks-mcp-server, currently 1.1.1, with binaries meshseeks and meshseeks-server. CLAUDE_CLI_PATH sets a custom path to the Claude CLI. On macOS the first run may fail while the OS asks for folder permissions.
One command — npx -y @thomaswalichiewicz/meshseeks-mcp-server