Auxiliary Intelligence · Tool use
Connect MCPs to your private model — not to Claude or GPT.
Wire up GitHub, Linear, web search, your own internal MCP servers — and your Auxen-deployed model gets the same tool access as a frontier-grade agent. Tool credentials are encrypted on Auxen's control plane. Reasoning runs on your single-tenant Auxen instance — no shared inference, no third-party routing — and conversation data is deleted when you destroy the instance.
How it works
- Go to Tools in your dashboard.
- Click Add Tool. Paste:
- Name — a label you'll recognize (e.g. “GitHub”)
- Endpoint URL — the MCP server URL (e.g.
https://mcp.github.com/mcp) - API key — optional; the Personal Access Token from the MCP provider
- Auxen validates the endpoint by calling
tools/list, encrypts your key (AES-256-GCM), and binds the tool to your instance. - From now on, every /api/chat request to that instance runs through an agentic loop — your model can use any of the registered tools during the conversation. You don't change your client code.
Where everything runs
- Customer app — sends a normal /api/chat request using its
auxk_…key. - Auxen control plane — injects tool schemas into the request, parses
tool_callsfrom the model, decrypts your stored credentials in memory, executes against the registered MCP endpoint, appends results, re-calls the model. Loops up to 10 iterations. - Vast.ai GPU — your dedicated Ollama instance. Inference only. No credentials. No tool logic.
- External MCP server — receives tool calls with your API key, executes against the real underlying service (GitHub, Linear, etc.), returns the result.
Limits
- 32 tool definitions per request (most MCPs fit)
- 10 iteration cap per chat request (covers complex agentic flows; aborts runaway loops)
- 5-second timeout per individual tool call (fails fast)
- One MCP per Auxen instance for MVP — multi-MCP support lands in v1.1
Security
- API keys encrypted at rest with AES-256-GCM. Master key lives in Auxen's env, never the DB.
- No tool arguments or tool results are logged in plaintext. Audit log records metadata only — tool name, latency, status, byte count.
- HTTPS-only transport to your registered MCP endpoints.
- You can disable or delete a tool binding at any time; deletion permanently removes the encrypted key.