Questions, answered straight
Is it actually free?
Yes. The core (agent-mesh, published on npm as meshfleet) is MIT licensed — spawning, messaging, routing, receipts, and councils are free forever. Pro adds signed ledgers and attestation reports on top; the boundary is “mechanisms free, assurance paid,” and mechanisms never move behind the paywall.
Does it phone home?
Not that we've found. Meshfleet runs on your machine, and the server's source contains no telemetry or analytics code. Its built-in clients connect only to loopback by default; if you choose to expose its SSE/A2A endpoint to other machines, it requires an auth token. The agent CLIs it launches (OpenCode, for example) talk to their own model providers, and fetching the package uses npm. The source is MIT, so you can check it; we haven't published a formal audit yet. There are no accounts or logins in the product. The waitlist on this website is separate: it stores the email you submit and sends us your note (see Privacy), and the site uses Cloudflare's cookieless Web Analytics.
Only for OpenCode?
OpenCode is where Meshfleet is native today — it's the runtime with the acute pain (serial agents, the 30-minute timeout). The core is a standard stdio MCP server, and the README ships configs for Claude Code, Codex, and any MCP client. Only the generic config is marked tested; Claude Code's follows its docs and is still awaiting verification. Spawned workers default to OpenCode's runner; other worker runtimes need operator setup. Broader host support tracks real demand: if you want it somewhere specific, say so on GitHub.
How is this different from LangGraph or CrewAI?
Those are Python frameworks you build an application around. Meshfleet is infrastructure for a coding session: TypeScript, local-first, native to your editor's agent runtime, and agents are real OS processes rather than nodes in a graph. Messaging is witnessed — receipts and quorum votes are built in, not bolted on.
What's the story behind it?
Meshfleet is a ground-up TypeScript port of the coordination bus that ran our own internal fleet — 10+ agents over roughly ten weeks, 18,404 messages. That was the builder's fleet, not a customer's. The receipts-and-councils design comes straight from what that fleet actually needed, not from a whiteboard.
How many agents can I run?
Agents are OS processes, so the practical limit is your machine and your model-provider rate limits — fleets of ten-plus are routine in testing. Start with three to five; you'll feel the parallelism immediately.
Is it safe to let agents message each other?
Safer than letting them share a context window. Messages are typed and logged, every delivery leaves a receipt, and consequential decisions can require a council vote instead of one agent's say-so. The full security posture is documented in the repo; report anything to security@meshfleet.app.