ScoreCompute

Connect it to your AI

Claude Code

claude mcp add --transport http scorecompute https://scorecompute.com/mcp

Claude, ChatGPT, Cursor and other apps

Add it as a custom connector with this URL (Streamable HTTP):

https://scorecompute.com/mcp

Cursor, VS Code and other apps (mcp.json)

{
  "mcpServers": {
    "scorecompute": {
      "url": "https://scorecompute.com/mcp"
    }
  }
}

Tools it offers

38 tools, as the server itself reported during our check on Oct 31, 2026.

• simulate_piEstimate pi with seeded Monte Carlo and a 95% Wilson interval; up to 5 million samples. CPU or optional CUDA. Other direct scientific engines run on CPU.
• optimize_tasksExactly maximize value for up to 32 independent, indivisible tasks within an integer-minute budget. Each task is selected at most once; no prerequisites or shared resources.
• verify_shadowsCheck measured shadows using NOAA solar position, a ground-plane homography from 4–16 controls and Monte Carlo uncertainty. Returns conditional consistency, never image authenticity.
• check_ephemerisCheck a claimed Sun (soleil) or Moon (lune) position for UTC time/location using NOAA/Meeus approximations and lunar parallax. Returns model-conditional angular consistency and lunar phase.
• audit_statisticsApply necessary GRIM/GRIMMER checks to rounded means and sample deviations of bounded integer observations. Passing does not prove a dataset exists; inconsistency does not establish fraud.
• analyze_kinematicsConvert angular speed and candidate distances to minimum transverse speeds; check a claimed speed-distance pair. Distance hypotheses are supplied, not measured; no object-origin conclusion.
• check_shadow_trackCompare 4–120 measured minute/azimuth pairs with solar motion: residuals, slope, direction and time-lapse speed. Supply measurements, not video. Results are model-conditional consistency findings.
• solve_equilibriumSolve Kuhn poker with CFR+ self-play, measure exploitability with pure best responses and compare the game value with -1/18. Illustrative game theory, not gambling advice.
• analyze_chess_gameCompare 4–40 supplied FEN positions and played SAN/UCI moves with Stockfish: first-choice agreement and centipawn loss. Entire analysis has a 25-second deadline; no cheating verdict.
• screen_wash_tradingScreen a supplied trade series for repeated volume at nearly equal prices within a declared time window. Return recycled-volume ratio and cycles; a pattern does not prove wash trading.
• simulate_planEstimate deadline success, mean duration and empirical p95 for up to 32 integer-minute tasks under independent bounded uniform duration variations. Seeded CPU sampling; not a real-world forecast guarantee.
• route_optimizerOrder 2–256 points from index 0 using planar or spherical kilometre distances. Exact only after completed Held–Karp search (at most 13 points); otherwise bounded heuristics. Returns every leg and stop reason. No roads or
• forge_route_policyTrain route-score ASTs on 1–8 supplied cases: at most 64 attempts and 31 nodes; stop between trials after 25s. Select converged candidates using training quality plus logical-fuel cost, a work proxy rather than CPU time.
• run_route_policyExecute a supplied bounded score AST on 2–128 Euclidean-km points with depot 0 and 2-opt. Validates the closed tour; exact=false. No retraining, external effects or promotion. Hashes identify programs, not independent va
• plan_golden_hourCompute geometric golden [-4,+6°] and blue [-6,-4°] windows for one UTC day, 1900–2100. Sunrise at -0.833° is a separate convention. No refraction, terrain or weather. Numerical tolerance is not astronomical accuracy; or
• triangulate_tracksFit position/velocity from 2–4 supplied azimuth/elevation tracks using constant velocity in spherical ECEF and declared angular noise/clocks. Reports fit status and conditional local uncertainty; no object identification
• plan_observationFit original tracks, then rank 1–64 supplied future series by local-linear uncertainty and declared cost. Assumes constant velocity, independent new errors and exact stations/clocks. Candidate angles and client baselines
• assimilate_observationRefit original tracks plus one supplied new angular series, each counted once; compute the baseline natively. Declare synthetic/measured origin (unverified). Conditional uncertainty may worsen; forecast angles are not ac
• stress_test_observationRefit 1–32 declared hypothetical angular/clock perturbations to challenge reconstruction stability. Explicit hypothetical acknowledgement required. Non-convergence is inconclusive; none_found covers examined cases only.
• improve_scheduleCompare up to 32 declared deadline/optional-task changes under independent-uniform durations; preserve required tasks. Search bounds select one finalist for separately seeded confirmation, without retries. Conditional on
• list_capabilitiesRead native contracts, themes, units, assumptions, backends and approved composition families; optional exact theme filter. tools/list discovers direct MCP tools; list_composition_capabilities adds planner bindings.
• plan_missionPreview robust scheduling without running engines: validate assumptions/permissions and return up to three candidate budgets (nominal, 10% reserve, 20% reserve), compute limits or blockers. No arbitrary tools or external
• build_robust_planCompare up to three task selections; choose highest examined value meeting the target under simultaneous 95% sampling bounds or an exact bound. Independent-uniform durations; can return inconclusive/blocked. Streams nati
• list_composition_capabilitiesRead implemented versioned contracts, native capabilities, approved adapters and planning examples. Unsupported unit/model facts expose missing adapters. No engines run; direct tools remain independently callable.
• compose_capabilitiesPlan from declared available/goal contracts within tool, locality and search bounds. Returns dependencies, blockers and existing-tool bindings. Cost counts graph steps, not speed, price or quality. Runs no engines; execu
• execute_compositionRun approved task-selection/route-to-duration risk chains only. Supply engine inputs and explicit uncertainty; route speed is 1–150 km/h. Returns native receipts and 95% sampling bounds. Completed means assessed, not saf
• network_statusRead contributor availability, offered CPU/GPU resources and verified throughput by workload. Hardware is self-reported; per-GPU VRAM is not pooled. Mixed-workload samples do not measure comparative speed.
• submit_network_jobQueue fixed public synthetic pi (CPU/CUDA) or one of three discrete schedule scenarios (CPU only). No private data/custom code. Returns a ticket; poll get_network_job. Coordinator recomputes every accepted shard; no net
• get_network_jobRead a public contributor ticket/progress and verified pi or discrete synthetic schedule result. Schedule results include fixed-sample Wilson95 bounds. Worker/verification/wall times are separate; model probabilities are
• check_arithmeticCheck an explicit arithmetic claim. Checked i128 rationals give exact verdicts when expression, claim and tolerances are representable. Unsupported functions or overflow return an estimate with an undetermined verdict. I
• compute_orbitCompute an ideal two-body Keplerian orbit around one solar mass. Time starts at periapsis; distances use AU and time uses days. This model is not a real celestial ephemeris.
• hear_the_shapeCompute 1–12 ideal fixed/free membrane modes for a simple polygon in metres using bounded P1 finite elements. Return upper bounds, extrapolated estimates and optional frequencies/shapes. Finite spectral agreement does no
• kepler_stabilitySurvey up to 512 initial orbits in a planar circular Sun–Jupiter model; at most 4 million reserved CPU steps. Return finite-window sensitivity, actual durations, sampled Jacobi drift and explicit stops. Bands do not prov
• propagate_uncertaintyCheck dimensions, then compare first-order GUM with seeded fixed-sample Monte Carlo for an explicit model of 1–16 inputs. Return uncertainty budgets, coverage intervals and convergence/linearization findings. Conditional
• scan_counterexampleBounded deterministic float64 scan of a scalar <= or >= claim over a declared grid of 1–6 variables. Reports numerical witnesses, undefined and uncertain evaluations, and the exact stop reason. A propagated screening gua
• shape_the_soundKac's question in reverse: design a polygonal drumhead whose lowest tones play requested frequency ratios (a fifth, a triad, a pentatonic scale), each accepted within ±tolerance_cents. proven_impossible is returned only
• tour_profileLocally normalized geometric descriptors of 3–256 planar points: distances, nearest-neighbour isolation, centroid spread, PCA collinearity and histogram. The principal-axis mirror residual is null for isotropic noncoinci
• verify_claimVerify a structured arithmetic, shadow, ephemeris, statistics or route claim with a bounded native engine. Returns a model-scoped verdict and portable Ed25519 server receipt; never truth, authenticity or fraud certificat

Questions