three.ws Brain

io.github.nirholas/brain-mcp

Documentation

@three-ws/brain-mcp

Any model, one interface — discover LLM providers and run chat completions through the three.ws multi-provider router, from any AI agent.

A Model Context Protocol server that gives any AI assistant the three.ws multi-provider LLM router over stdio. Discover which models are available — Claude, the GPT-5.6 family, o3, Grok, Qwen, DeepSeek, NVIDIA Nemotron, IBM Granite, and more — and run a chat completion through whichever one fits, without wiring each vendor SDK yourself.

The router fronts every provider behind one endpoint and transparently falls back across mirrors and free tiers if the chosen model is briefly down. The free open-weight tiers (GPT-OSS 120B, NVIDIA NIM models) work with no key; the paid first-party flagships unlock with a three.ws API key.

Install

npm install @three-ws/brain-mcp

Or run with npx (no install):

npx @three-ws/brain-mcp

Quick start

Claude Code, one line:

claude mcp add brain -- npx -y @three-ws/brain-mcp

To unlock the paid flagships, pass your three.ws API key:

claude mcp add brain -e THREE_WS_API_KEY=sk_live_… -- npx -y @three-ws/brain-mcp

Claude Desktop / Cursor (claude_desktop_config.json or mcp.json):

{
	"mcpServers": {
		"brain": {
			"command": "npx",
			"args": ["-y", "@three-ws/brain-mcp"],
			"env": { "THREE_WS_API_KEY": "sk_live_…" }
		}
	}
}

Inspect the surface with the MCP Inspector:

npx -y @modelcontextprotocol/inspector npx @three-ws/brain-mcp

Tools

Tool	Type	What it does
`list_providers`	read-only	List every LLM the router can reach — key, network, tier, max output, live availability, and whether it needs auth.
`chat`	read-only	Run a chat completion through the chosen provider; returns the full reply, the route taken, token usage, and timing.

Both tools read live data — the provider set and completions vary between calls, so neither is idempotent. chat does not mutate any platform state, so it is annotated read-only.

Input parameters

list_providers — no parameters.

chat — messages (required: 1–100 turns of { role: "user" | "assistant", content }, content ≤ 16,000 chars), provider (provider key from list_providers, default gpt-oss-120b), system (optional system prompt, ≤ 8,000 chars), maxTokens (optional, 64–16384, default 4096, clamped to the model's ceiling).

Example

Both responses below are shape illustrations — the live provider set and generations will differ.

// list_providers (example response)
> {}
{
  "ok": true,
  "count": 18,
  "available": 11,
  "default_provider": "gpt-oss-120b",
  "providers": [
    { "key": "gpt-oss-120b", "label": "GPT-OSS 120B", "network": "OpenAI · OpenRouter", "tier": "balanced", "maxOutput": 8192, "available": true, "requiresAuth": false },
    { "key": "claude-sonnet-4-6", "label": "Claude Sonnet 4.6", "network": "Anthropic", "tier": "balanced", "maxOutput": 16384, "available": true, "requiresAuth": true },
    { "key": "deepseek-r1", "label": "DeepSeek R1", "network": "DeepSeek", "tier": "reasoning", "maxOutput": 8192, "available": true, "requiresAuth": true }
    /* … */
  ]
}

// chat (example response)
> {
    "provider": "claude-sonnet-4-6",
    "system": "You are terse.",
    "messages": [{ "role": "user", "content": "Name three Solana RPC methods." }]
  }
{
  "ok": true,
  "provider": "claude-sonnet-4-6",
  "model": "Claude Sonnet 4.6",
  "network": "Anthropic",
  "tier": "balanced",
  "routed_via": null,
  "content": "getBalance, getAccountInfo, getSignaturesForAddress.",
  "usage": { "input_tokens": 24, "output_tokens": 14, "total_tokens": 38 },
  "timing_ms": { "first_token": 412, "total": 980 }
}

routed_via is null when the requested provider answered directly; otherwise it names the mirror or free-tier route the router fell back to (so you always get an answer even if the first choice is briefly down).

Model availability is dynamic. Always call list_providers first rather than hardcoding a key — the current Claude ids are claude-fable-5, claude-opus-4-7, claude-sonnet-4-6, and claude-haiku-4-5, but the live list is the source of truth.

Authentication

Model class	Needs a key?
Free open-weight tiers (GPT-OSS 120B, NVIDIA NIM: Nemotron / Kimi / Llama 4)	No
Paid first-party flagships (Claude, GPT-5.6, o3, Qwen Plus, DeepSeek, …)	Yes — set `THREE_WS_API_KEY`

Create an API key at three.ws/account. Without it, paid models return a sign-in error and list_providers flags them with requiresAuth: true.

Requirements

Node.js >= 20.
Network access to https://three.ws (or your own THREE_WS_BASE).

Environment variables

Variable	Required	Default	Purpose
`THREE_WS_BASE`	no	`https://three.ws`	API base URL (override to self-host or target a preview).
`THREE_WS_API_KEY`	no	—	Bearer credential that unlocks the paid first-party flagships.
`THREE_WS_TIMEOUT_MS`	no	`120000`	Per-request timeout (completions stream server-side).

Errors

A failed tool call returns an MCP error result (isError: true) whose text is a single JSON object — { "ok": false, "error": "<code>", "message": "…" }, plus status and detail on upstream rejections:

`error`	Meaning	Recovery
`upstream_error`	The router rejected the request — `status` carries the real HTTP code: `400` (bad arguments or an unknown provider key), `401` (paid model without a valid `THREE_WS_API_KEY`), `429` (rate-limited), `502` (the model's stream failed before producing output).	Act on `status`; for `401` set a key, for an unknown key re-run `list_providers`.
`timeout`	No complete response within `THREE_WS_TIMEOUT_MS`. Long flagship generations can be slow — the default 120 s matches the router's own ceiling.	Retry, lower `maxTokens`, or pick a `fast`-tier provider.
`network_error`	The request never reached the router (DNS, offline, TLS).	Check connectivity / `THREE_WS_BASE`.
`bad_config`	`THREE_WS_TIMEOUT_MS` is not a positive number (thrown at startup).	Fix the env var.

If the requested provider fails mid-flight the router transparently falls back to a mirror or free tier and the reply's routed_via names the route taken — you only see an error when every route is exhausted.