npm
Sign UpSign In

@tensorfeed/mcp-server
TypeScript icon, indicating that this package has built-in type declarations

2.0.0 • Public • Published

TensorFeed MCP Server

The MCP server has its own repo: https://github.com/RipperMercs/tensorfeed-mcp

User-facing docs, install instructions, and the full tool reference live there. Star and watch that repo to follow MCP server updates.

This subfolder remains in the main tensorfeed repo as the publishing source for the npm package (@tensorfeed/mcp-server) and the official MCP registry entry. Edits to src/, server.json, package.json, etc. happen here and get pushed to the standalone repo on release.

Featured: Route Verdict

The single best model to use right now, as one signed call. route_verdict fuses live pricing, contamination-discounted benchmark capability, real production usage, measured p95 latency probes, live incident state, and deprecation flags into one ranked decision, with an AFTA-signed receipt over the exact inputs. Instead of stitching together pricing pages, benchmark leaderboards, status dashboards, and your own latency tests, you get a current, defensible routing answer in one request.

Zero install, no key, one command (works today)

curl -s -A "tensorfeed-cc-quickstart" "https://tensorfeed.ai/api/preview/route-verdict?task=code"

Swap task for reasoning, creative, or general, or pass ?model=<id-or-name> to score a specific model. The free preview is 10 calls per day per IP, no token. Abridged real response:

{
  "ok": true,
  "preview": true,
  "query": { "task": "code", "model": null },
  "verdict": {
    "rank": 1,
    "model": { "name": "Gemini 2.5 Pro", "provider": "google" },
    "pricing": { "blended": 5.625, "unit": "per 1M tokens" },
    "quality": { "trust_discounted": 0.6498 },
    "latency": { "measured_p95_ms": 1223, "source": "measured_probe" },
    "operational": { "ok": true, "status": "operational" },
    "composite_score": 0.8449,
    "why": "code quality 0.6498 after trust discount; corroborated by real usage (rank 5, 6.5% share, flat); measured p95 1223 ms; operational; blended $5.625 / 1M"
  },
  "rate_limit": { "limit": 10, "remaining": 9, "scope": "per IP per UTC day" },
  "upgrade": {
    "premium_endpoint": "/api/premium/route-verdict",
    "adds": ["runners_up", "AFTA-signed receipt", "filter params", "no rate limit"]
  }
}

The agent path (MCP)

With @tensorfeed/mcp-server installed, an agent gets one route_verdict tool with a tier parameter. Call it with the default free tier for the pick, then tier="full" when it needs to defend the choice:

# Free taste: the top pick + reasoning, no token (10/IP/day). tier defaults to "preview".
route_verdict({ task: "code" })

# 1 credit: ranked runners-up, constraint filters, AFTA-signed receipt
route_verdict({ task: "code", tier: "full", max_latency_p95_ms: 1500, budget: 8, min_quality: 0.6 })

tier="full" adds the ranked runners-up, the constraint filters (max_latency_p95_ms, budget, min_quality, require_operational, exclude_deprecated), and the AFTA-signed receipt the agent can audit later. Credits come from tensorfeed.ai/developers/agent-payments.

Why it matters

Models, prices, and latency move week to week. route_verdict is one signed call an agent can act on now and later prove why it routed the way it did, without rebuilding the comparison from scratch each time.

Catalog

24 tools. The core flagships, the eight signed verdicts, the time-series tools, and the webhook watches are dedicated tools; the rest of the 100+ TensorFeed endpoints are reachable through the find_tensorfeed_data discovery tool and callable over HTTP. Free tiers need no token; paid tiers charge USDC on Base via x402 and return an AFTA-signed receipt. The full tool reference lives in the standalone repo.

Verdict family

Eight signed decisions (route_verdict is featured above). Each is a single tool with a tier parameter: tier="preview" (default) is free, tier="full" costs 1 credit ($0.02) and adds the full ranking and an AFTA-signed receipt:

  • provider_reliability_verdict: the safest AI provider to build on, ranked by availability and tail consistency over TensorFeed's own probes.
  • x402_settlement_verdict: the x402 settlement momentum, concentration, and leading publisher over a 24h, 7d, or 30d window.
  • x402_publisher_verdict: a signed trust verdict on one publisher domain before you pay it.
  • stack_safety_verdict: a GO, HOLD, or BLOCK deploy gate over your package@version pins, with the worst CVE or KEV match.
  • benchmark_trust_verdict: a trust band and 0 to 100 score for a benchmark, flagging saturation, contamination, and held-out status.
  • failover_verdict: when a provider is degraded, the single best operational provider to fail over to, with ranked alternatives.
  • ssvc_verdict: the CISA SSVC Act, Attend, Track, or Track* decision for one CVE, with a live KEV cross-check.

Publish a new version

From the main tensorfeed repo:

# 1. Bump the version in mcp-server/package.json + mcp-server/server.json
# 2. Build + npm publish from the mcp-server/ folder
cd mcp-server
npm run build
npm publish --access public

# 3. Republish to the official MCP registry. The script lives at
#    repo-root/scripts/, not mcp-server/scripts/, so step back up first.
cd ..
.\scripts\mcp-publish.ps1

# 4. Mirror to the standalone repo - automated. The
#    .github/workflows/mirror-mcp-server.yml workflow runs on every
#    push to main that touches mcp-server/. To trigger a manual sync,
#    go to the Actions tab and run "Mirror MCP server to standalone
#    repo" via workflow_dispatch.

Automated mirror setup (one-time)

The mirror workflow needs a personal access token with contents: write permission on RipperMercs/tensorfeed-mcp. Set it once:

  1. Generate a fine-grained PAT at github.com/settings/personal-access-tokens scoped to the standalone repo only, with Repository permissions Contents: Read and write and Metadata: Read-only.
  2. Add it to this monorepo as a repository secret named STANDALONE_REPO_TOKEN (Settings -> Secrets and variables -> Actions).
  3. The workflow will pick it up on the next push to main that changes mcp-server/**, or via the manual "Run workflow" button.

Quick links

Readme

Keywords