Self-hosted retrieval router for AI agents: budgets, provider routing, route receipts.
One search endpoint for AI agents. Put a router in front of Brave, Tavily, Exa, Firecrawl, and Crossref. Every request carries a hard cost ceiling and deadline; every response comes back with a receipt that says which provider ran, why, and what it cost.
Your agents stop hard-coding a provider into prompts and app code. They send one request shape; the gateway picks an eligible provider under a policy you control, enforces the budget and deadline before the call, retries safely on failure, and returns normalized results plus a full trace. You keep your provider keys — they never leave your gateway.
Nothing is a black box. Every /v1/search response carries the routing decision:
"route": {
"selectedProvider": "brave",
"attemptedProviders": ["brave"],
"reasonCodes": ["TASK_MATCH", "WITHIN_BUDGET", "RELIABILITY_WEIGHT"],
"attempts": [
{ "provider": "brave", "outcome": "success", "latencyMs": 640, "estimatedCostUsd": 0.005 }
],
"estimatedCostUsd": 0.005,
"latencyMs": 640,
"fallbackUsed": false,
"traceId": "rt_b400e7c8"
}
agent ──▶ { query · task · maxCostUsd · maxLatencyMs }
│
▼
┌──────────────┐ your keys
│ FetchMux │ ─────▶ Brave · Tavily · Exa
│ policy │ Firecrawl · Crossref
└──────────────┘ ◀───── (bring your own)
│
▼
agent ◀── evidence[] + route receipt
A provider is eligible only when its credentials, task fit, circuit state, spend, and deadline all pass. Budgets and deadlines are eligibility rules, not best-effort hints. Fallback happens only on retryable failures.
No provider account needed — the public Crossref route runs out of the box:
git clone https://github.com/krutftw/fetchmux
cd fetchmux
npm install
npm run build
export FETCHMUX_API_KEY="a-long-random-key"
export CROSSREF_ENABLED=true
export CROSSREF_CONTACT_EMAIL="you@example.com"
npm run dev:gateway
From another shell:
curl http://127.0.0.1:8787/v1/search \
-H "Authorization: Bearer a-long-random-key" \
-H "Content-Type: application/json" \
-d '{ "query": "retrieval augmented generation", "task": "scholarly", "maxLatencyMs": 8000 }'
To route real web search, set a provider key and use a web task instead:
export FETCHMUX_API_KEY="a-long-random-key"
export BRAVE_API_KEY="your-brave-key"
export BRAVE_COST_PER_REQUEST_USD="0.005" # from your provider plan
npm run dev:gateway
New to Firecrawl? New accounts get 10% off the first month through this link (referral — FetchMux earns a small commission, no extra cost to you).
Point any MCP client (Claude, Cursor, and friends) at the published server:
{
"mcpServers": {
"fetchmux": {
"command": "npx",
"args": ["-y", "@fetchmux/mcp"],
"env": {
"FETCHMUX_BASE_URL": "http://127.0.0.1:8787/",
"FETCHMUX_API_KEY": "your-gateway-key"
}
}
}
}
Two read-only tools: search_web and preview_search_route.
Or use the typed SDK, @fetchmux/sdk:
import { FetchMux } from "@fetchmux/sdk";
const client = new FetchMux({
baseUrl: "http://127.0.0.1:8787/",
apiKey: process.env.FETCHMUX_API_KEY,
fetch: globalThis.fetch.bind(globalThis),
});
const res = await client.search({
query: "latest stable Node.js release",
task: "fresh_facts",
maxCostUsd: 0.02,
});
Bring your own key for each. Set the matching *_API_KEY, plus an optional
*_COST_PER_REQUEST_USD if you want dollar budgets enforced.
| Provider | Use | Key |
|---|---|---|
| Brave | web search | BRAVE_API_KEY |
| Tavily | web search, research | TAVILY_API_KEY |
| Exa | web search, docs | EXA_API_KEY |
| Firecrawl | page content | FIRECRAWL_API_KEY |
| Crossref | scholarly metadata | none (CROSSREF_ENABLED=true) |
| Method | Path | Auth | Behavior |
|---|---|---|---|
GET | /health | public | Process health and version |
GET | /ready | public | Provider readiness |
GET | /v1/providers | bearer | Provider configuration status |
POST | /v1/route/preview | bearer | Ranked candidates, no provider call |
POST | /v1/search | bearer | Routed retrieval and route receipt |
Full contract: docs/openapi.yaml.
The process does not auto-load .env in local Node development; set variables in the shell or a
process manager. Docker Compose reads the ignored .env file.
| Variable | Default | Purpose |
|---|---|---|
FETCHMUX_API_KEY | none | Protected-route bearer key |
FETCHMUX_API_KEYS | none | Comma-separated keys for rotation |
FETCHMUX_AUTH_DISABLED | false | Exact true bypasses auth (trusted local use only) |
FETCHMUX_ALLOWED_ORIGINS | none | Comma-separated browser origins; no CORS when empty |
FETCHMUX_HOST | 127.0.0.1 | Bind address |
FETCHMUX_PORT | 8787 | TCP port |
BRAVE_API_KEY / TAVILY_API_KEY / EXA_API_KEY / FIRECRAWL_API_KEY | none | Provider credentials |
CROSSREF_ENABLED | false | Exact true enables the credential-free scholarly route |
CROSSREF_CONTACT_EMAIL | none | Monitored contact for Crossref's polite pool |
*_COST_PER_REQUEST_USD | none | Per-provider cost estimates used by dollar budgets |
See provider configuration before enabling maxCostUsd.
cp .env.example .env # add your keys, never commit it
docker compose up --build -d
curl http://127.0.0.1:8787/health
Non-root Distroless image: Linux capabilities dropped, read-only root filesystem, provider credentials passed only at container start.
Validate every case and provider pairing with no network calls or credits:
npm run benchmark -- --workload benchmarks/workloads/founding-v1.json --mode dry-run
Live mode needs provider keys and an explicit --confirm-live. Check each provider's terms before
publishing results — see the benchmark methodology.
Open source, self-hosted, single-tenant, BYOK. Route events go to stdout as JSON and exclude your query text, keys, and result content by default. No database, no telemetry.
It is not a hosted service, a pooled-credit reseller, or a claim that these providers are interchangeable. Provider names are the adapters it ships with, not partnerships. A hosted version is on the roadmap — star the repo to follow.
npm test # 232 tests
npm run typecheck
npm run lint
npm run build
npm run dev:gateway
npm run dev:site
More docs: product design · local development · deployment · provider configuration · data handling · incident response
Security issues: security@fetchmux.com.
Apache-2.0. Free to self-host, modify, and redistribute.
Source-derived launch command. Check the maintainer’s required arguments and credentials before running:
npx -y @fetchmux/mcpMerge this template into ~/Library/Application Support/Claude/claude_desktop_config.json. Keep existing servers. Add any arguments, credentials, and permissions required by the maintainer; this template has not been install-tested.
{
"mcpServers": {
"io-github-krutftw-fetchmux": {
"command": "npx",
"args": [
"-y",
"@fetchmux/mcp"
]
}
}
}Restart Claude Desktop completely for changes to take effect. Confirm the server appears connected in the client’s tool list, then try a read-only example from its documentation.
Claude Desktop setup referenceio.github.krutftw/fetchmux works with any MCP-compatible client. Copy the config snippet from the Configuration section above and add it to the file shown for your client, then restart the application.
~/Library/Application Support/Claude/claude_desktop_config.jsonRestart Claude Desktop completely for changes to take effect.~/.cursor/mcp.jsonRestart Cursor for changes to take effect..vscode/mcp.jsonReload VS Code window for changes to take effect.~/.codeium/windsurf/mcp_config.jsonRestart Windsurf for changes to take effect..mcp.jsonSave at the project root, then start Claude Code in that project and review the MCP server approval prompt. Keep real credentials out of shared files.