VulpineOS exposes a 36-tool MCP toolbelt that any model client can drive — Claude, GPT, Gemini, open-weights models, hosted aggregators, local runners. Bring your own keys; VulpineOS doesn't proxy your LLM traffic.
Every VulpineOS agent operates a Camoufox browser session through a Model Context Protocol server. The server exposes 36 typed tools: navigation, click, type, snapshot, find, verify, page-settled, fill-form, select-option, human-realism inputs, scripting, and more.
Any LLM client that speaks MCP can drive it. For clients that don't speak MCP natively, the Foxbridge translation layer exposes the same surface over Chrome DevTools Protocol.
The list is curated, not exhaustive — VulpineOS ships with 30+ providers pre-configured. Anything OpenAI-compatible can be wired up in seconds.
The headline frontier-lab models, end to end.
Open-source models served by their original labs or hosts.
Speed-first hosts for latency-sensitive agent loops.
Routers and marketplaces that fan out across many backends.
Run models on your own hardware. No API key required.
VulpineOS doesn't sit between you and the model provider. You configure your own API keys in the runtime config, and the agent talks to the provider directly. We don't see, log, or charge for tokens — your billing is between you and the provider.
This is a deliberate choice. Hosted runtimes that proxy LLM calls add a third party to your data path. The VulpineOS open-source runtime keeps that path local, even when running fleets of hundreds of agents.
Configuration lives in ~/.vulpineos/config.jsonor the panel's settings tab. Pick a provider, paste a key, choose a model. Every agent spawned afterward uses the configured stack. Per-agent overrides are supported when you want a specific model for a specific task.
The first-launch wizard walks through provider selection automatically — the runtime won't spawn agents until at least one provider is configured.
docs.vulpineos.com — full list of supported providers and configuration paths.
docs.vulpineos.com/mcp-tools — every tool, every parameter.
/integrations/openclaw — the bundled agent loop that drives the toolbelt by default.
Self-host the runtime and point your model client at the bundled MCP server. No vendor lock-in, no proxy, no token surcharge.
