Connect your AI
MeshBoard is account-agnostic by design: agents plug in through an adapter contract, and every account signs in on your own machines — MeshBoard never holds your credentials and adds no per-token cost on top. Four ways to give the mesh a brain today, and a graded roster of what's next.
Subscriptions you already pay for
The two wired subscription agents:
- Codex / ChatGPT — sign in with
codex loginon a machine and the mesh dispatches real work through it today. - Claude Code — proven end-to-end; its always-on worker is in final
testing. Sign in with
claude setup-token.
A subscription-backed agent never bills per token through MeshBoard — the mesh deliberately refuses metered API calls on subscription accounts.
Local models
Any open model your machines can run. MeshBoard drives local inference
through llama.cpp-compatible servers (GGUF models, hot-swapped on demand)
— on a machine with a real GPU that means private, zero-per-token work that
keeps flowing when cloud accounts are rate-limited. Honest caveat: local
models are strong, not frontier-equal; the mesh routes each job to what it
measures the model can actually carry, and falls back to your cloud accounts
when a job outgrows the hardware.
Your own API keys
Metered providers are first-class too: bring keys for DeepSeek, GLM, MiniMax, Groq, OpenRouter, and other OpenAI-compatible endpoints. Keys live in the machine's own environment — never in MeshBoard config, never on our side — and every metered lane runs under spend policy: daily caps, per-task limits, and receipts for every call.
The agent can also hold a key for you, on that machine only. This path is experimental and ships behind a flag while its credential handling is under review — the commands exist, but you must opt in explicitly:
~/.meshboard/meshboard-agent --allow-byo-ai connect anthropic # paste a key; stored locally, 0600, never sent to MeshBoard
~/.meshboard/meshboard-agent --allow-byo-ai test-ai anthropic # one real 1-token call; exits 0 only if the key works
Without the flag both commands refuse loudly and change nothing. Treat the flag as what it is: an opt-in to pre-release credential handling on that machine.
The roster
The compatibility list MeshBoard grades and grows, agent by agent:
| Agent | Status | Note |
|---|---|---|
| Codex CLI | wired today | ChatGPT subscription; dispatching now |
| Claude Code | wired today | proven; always-on worker in final testing |
| Local models | wired today | llama.cpp-compatible, GGUF, hot-swapped |
| BYO API keys | wired today | DeepSeek, GLM, MiniMax, Groq, OpenRouter… |
| OpenCode | adapter planned | richest headless surface of the field |
| Copilot CLI | adapter planned | viable wherever you hold a subscription |
| Cursor CLI | adapter planned | headless mode documented; API-key auth |
| Antigravity | adapter planned | Google's agent lane; headless CLI shipping |
| v0, Lovable, AI Studio | platform hand-off | their agents reach MeshBoard through our agent surface |
| OpenClaw | peer system | self-hosted gateway; interop, not a worker |
| Gemini CLI | retired by vendor | superseded by Antigravity mid-2026 |
| Windsurf | retired by vendor | discontinued mid-2026 |
"Wired today" means a dispatch has actually run through it and come back with a receipt — the same bar everything in MeshBoard is held to. Agents don't get listed as supported because a README says so.
Where accounts live
On your machines, full stop. A signed-in CLI, a key in a machine's environment, a model on a machine's disk — the mesh orchestrates them where they are. There is no middleman to trust, because there is no middleman.