Meridian MCP

Dynamic task routing via orbital mechanics. Domain-agnostic — candidates can be tools, prompts, documents, products, or any routable entity.

Per request, an LLM (Llama-3.3-70B via GitHub Models' free tier) emits N candidate routing entries. A deterministic classifier extracts a 9-scalar physics signature from each candidate's content alone — no curated lookup: mass (log-scaled body length × keyword count), scope, independence, cross_domain affinity (token-domain entropy across three star systems), fragmentation, drag, dep_ratio (max sibling Jaccard), lagrange_potential, coherence_time (g⁽¹⁾-style autocorrelation over the candidate's token stream — added in 3.1.0), plus orbital and optical parameters (semi-major axis, eccentricity, inclination, period, perihelion, aphelion, mean anomaly; wavelength, polarization, amplitude, phase). Six per-class scoring rules assign a celestial body class by argmax: planet, moon, trojan, asteroid, comet, or irregular.

The class-scoring rules:

score_planet    = min(mass, scope, independence)^1.5
score_moon      = 2 · max(0, ½ - independence) · 𝟙[parent] · (1 - mass/2)
score_trojan    = dep_ratio · 𝟙[parent] · (1 - fragmentation)
score_asteroid  = 2.5 · max(0, 0.55 - mass) · scope · independence
score_comet     = drag · cross_domain · (1 - dep_ratio)
score_irregular = 0.85 · cross_domain · fragmentation

class(p) = argmax_c score_c(p)

Output: a deterministically ranked list with route_score, full classification, and decision rule per candidate. Wire format: Model Context Protocol over stdio or Streamable HTTP. Stdio shim is ~5 KB. Tested with Claude Code, Cursor, Windsurf, Goose, Continue, Grok custom connectors, ChatGPT custom MCPs, and Claude.ai connectors.

3.0 — renamed from meridian-skills-mcp. The classifier was always domain-agnostic; the "skills" framing biased the LLM prompt toward AI-agent capabilities. v3 drops that framing across the prompt, code, branding, and npm name. Migration: npm i -g meridian-orbital (the old package is deprecated; both binaries are still named meridian-mcp / meridian-mcp-http so client configs keep working). The hosted HTTP MCP at mcp.ask-meridian.uk/mcp continues to work — URL unchanged.

Install (stdio — Claude Code / Cursor / Windsurf)

npm install -g meridian-orbital
claude mcp add meridian meridian-mcp

Same install works in Cursor, Windsurf, Goose, Continue, and any MCP client that speaks stdio.

You'll need a GitHub personal access token with the Models: read permission (free tier). Generate one at https://github.com/settings/personal-access-tokens/new and export it:

export MERIDIAN_GITHUB_TOKEN=github_pat_...

(The MCP also picks up plain GITHUB_TOKEN if you have one already in your environment.)

Use as a Grok connector

A hosted Streamable-HTTP variant lives at https://mcp.ask-meridian.uk/mcp with full OAuth 2.1 + PKCE so it slots into any host that requires a connector URL — Grok's custom MCP connectors, ChatGPT custom MCPs, Claude.ai connectors. No npm install, no PAT entry from your side, no infra.

In Grok's "Add custom connector" dialog, paste these:

When you click "Authorize" in Grok, it opens /authorize — a one-click confirmation page (no PAT pasting, no GitHub jargon). Inference runs against GitHub Models using the operator's PAT, so end users see zero friction. Tokens last 1 hour and can be reauthorized any time.

The same URL works for ChatGPT custom MCPs and Claude.ai connectors — they speak the same MCP Streamable HTTP + OAuth 2.1 spec.

Self-hosting the HTTP variant

If you'd rather operate your own remote MCP, the package ships a Node binary:

npx -y meridian-orbital meridian-mcp-http
# → listening on http://0.0.0.0:3333/mcp · auth=pass-through · v3.1.0

Or via Docker (MCP_MODE=http flips the entrypoint):

docker run --rm -p 3333:3333 -e MCP_MODE=http meridian-orbital

Auth modes:

• Pass-through (default). Each call's Authorization: Bearer … is forwarded to GitHub Models. Users bring their own PAT.
• Shared gateway. Set MERIDIAN_GATEWAY_TOKEN (what callers pass) + MERIDIAN_GITHUB_TOKEN (what the server uses for inference).

The hosted Worker variant additionally implements OAuth 2.1 + PKCE; the Node binary is bearer-only (suitable for stdio→HTTP bridges and tools like curl).

What it does

Single tool: route_task(task, limit?).

input: a natural-language task
   ↓
GitHub Models (Llama-3.3-70B) generates 5 candidates
   ↓
local orbital classifier
   • derives physics: mass, scope, independence,
     cross_domain, fragmentation, drag, dep_ratio
   • assigns class: planet | moon | trojan |
                    asteroid | comet | irregular
   • computes star-system membership (forge / signal / mind),
     parent candidate, Lagrange potential
   ↓
output: ranked candidates with full bodies, classifications,
        and decision rules

Typical call takes 5–15 seconds. Each result ships its full markdown body so the caller agent can lift the candidate straight into its context window.

Configuration

What changed in 2.0.0

The 1.x line called a Cloudflare Worker (https://ask-meridian.uk/api/orbital-route) that ran the LLM and orbital classifier server-side. That backend has been retired. 2.0.0:

• Self-contained. The orbital classifier runs in-process. The LLM call goes to GitHub Models directly. No backend dependency.
• Bring-your-own token. Free GitHub tier, generous quota.
• Faster. 5–15 s instead of 30–50 s (no extra network hop, GitHub's inference is quick).
• Same output shape. Drop-in replacement; no agent prompt changes needed.

To keep using the closed-domain Python scorer + curated 88-entry corpus that shipped with 0.3.x, pin to meridian-skills-mcp@0.3.2. To keep calling the now-defunct Cloudflare backend, pin to 1.0.1 (will fail with HTTP 405 on every call).

Web miniapp + the live remote MCP

Same orbital classifier powers two front-ends served from mcp.ask-meridian.uk:

• ask-meridian.uk/miniapp — type a task, see the candidates orbit. Calls the live MCP at mcp.ask-meridian.uk/v1/route, same Llama-3.3-70B + classifier path the connector uses.
• meridian.ask-meridian.uk/lens/ — WebXR Vision Lab. Captured headset frames POST to mcp.ask-meridian.uk/v1/vision (GPT-4o-mini, operator-paid), candidates orbit anchored star systems in-view. Same backend as miniapp.

Both call the first-party browser endpoint /v1/route — Origin-allowlisted, operator-paid, no PAT pasting. The OAuth-gated /mcp endpoint (this section's "Use as a Grok connector" path) is unchanged.

Online learning loop

The browser endpoint /v1/route applies a fitted-correction layer on top of the heuristic ranking. Every time a user engages a candidate (planet click in lens, detail-panel open in miniapp, card click in vision-lab), the front-end POSTs to /v1/feedback and the worker runs one pairwise-ranking SGD step against the chosen candidate vs every other. Constant per-request cost (~1 ms), no GPU, no local execution.

user click → /v1/feedback → KV → SGD step → updated weights → next /v1/route uses them

• final_score = heuristic_route_score × (1 + tanh(K · w·x)) — bounded to [0, 2], so no individual candidate can be silently boosted beyond 2× heuristic.
• 25-feature vector per candidate: 9 physics scalars (the 8 originals plus coherence_time added in 3.1.0) + 6 class one-hot + 3 star-system one-hot + 3 token-hit features + 4 ranking features. Stored under FEATURE_VERSION=v2 in KV; bumping the version re-inits weights cleanly.
• Cold start: w = 0, multiplier = 1, pure heuristic. Day 1 deployments don't need any training data.
• The OAuth-gated /mcp path (Grok / ChatGPT / Claude.ai connectors) keeps deterministic heuristic ranking for reproducibility.
• Two GitHub Actions cron jobs close the loop without organic traffic: classifier-bootstrap.yml (every 3 days, feeds labelled examples from a public HF benchmark into /v1/feedback) and classifier-health.yml (Mondays, posts recall@1 / @5 + model state to landing/healthz.json).

Read-only model state: GET https://mcp.ask-meridian.uk/v1/model-info.

Full architecture + the calibration journey that produced this design (the planet-bias bug, the two textbook physics frameworks we tried and abandoned, the v2 retune, the 81% recall@1 [95% Wilson CI 60%, 92%] finding on real labelled data): blog post.

License

MIT — see LICENSE.