One endpoint.
Independent tools.

Public MCP endpoint

https://scorecompute.com/mcp

Choose a client that supports remote MCP servers over Streamable HTTP. Public access requires no API key. Configuration field names vary by client; a URL-based example is:

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

Protocol sequence

1. Initialize the MCP connection with the software client name and version.
2. Use tools/list to discover 38 tools and their current input schemas: 22 scientific tools, 6 Discovery tools, three contributor-network tools, six optional composition tools and Verify.
3. Call a tool with tools/call, preserving units and parameter bounds.
4. Read the structured result, assumptions and provenance. Close the session when finished.

This is a stateful endpoint. Clients retain the Mcp-Session-Id returned by initialization. Sessions expire after inactivity; reinitialize after a session-expired response.

What does “bounded” mean?

Every tool has a fixed maximum input size and amount of work, stated on its engine page. A request outside those limits is refused, and a long calculation ends with an explicit status or a timeout instead of running indefinitely.

Which ScoreCompute tool runs Monte Carlo on my own model?

Use propagate_uncertainty for a measurement model and simulate_plan for uncertain task durations. simulate_pi only estimates π.

Direct tools stay independent

Call any of the scientific or Discovery tools directly. For example, ask optimize_tasks to select a high-value schedule, then pass its selected durations to simulate_plan if you want to choose each step yourself. No orchestrator is required to use an engine.

Optional native mission composition

list_capabilities describes contracts, units, assumptions, execution backends and composition readiness, with an optional exact theme filter. This native catalogue includes the direct engine contracts; inclusion does not imply a planner binding. The hear_the_shape, shape_the_sound, propagate_uncertainty and kepler_stability engines are directly callable, with orchestration disabled. The separate typed capability catalogue also includes triangulation and four bound Discovery study tools; the two Method Lab tools are independently callable and not registered as composition bindings. The scheduling family below remains available; geometric route assessment uses the typed composition tools. plan_mission validates a request and previews its candidate budgets and work limits without running engines.

build_robust_plan compares up to three task selections using nominal, 10% and 20% time reserves. Native Rust validates the integer-minute handoff, uses exact worst-case bounds when possible, and otherwise runs fixed seeded Monte Carlo. Identical selections reuse their evaluation within that mission. This is one composite MCP call: its internal engines execute natively, without fabricated nested MCP requests.

{
  "name": "build_robust_plan",
  "arguments": {
    "items": [
      {
        "name": "A",
        "duration": 60,
        "value": 90
      },
      {
        "name": "B",
        "duration": 40,
        "value": 70
      },
      {
        "name": "C",
        "duration": 20,
        "value": 20
      }
    ],
    "budget": 100,
    "uncertainty": 0.25,
    "min_success_rate": 0.95,
    "samples": 10000,
    "seed": 42,
    "max_candidates": 3,
    "duration_model": "independent_uniform",
    "allowed_tools": [
      "optimize_tasks",
      "simulate_plan"
    ]
  }
}

Required inputs are tasks, budget and uncertainty. The other fields above show the defaults. Budgets and durations use integer minutes. Missions accept up to 32 tasks, a 1–1440 minute budget, uncertainty from 0 to 0.5, 100–20000 samples per candidate and at most three candidates. Unsupported duration models and missing tool permissions return explicit blockers; no external commands or destinations are accepted.

The selected candidate has the highest value among examined candidates whose simultaneous 95% Hoeffding lower bound meets the success target under the declared independent uniform duration model. If no candidate qualifies, the result is inconclusive with selected: null. An exact bound can replace simulation; in that case simulation: null is intentional. These bounds describe the supplied model, not real-world delivery guarantees or a globally optimal robust schedule.

Actual native events stream through MCP progress messages prefixed SC_MISSION_EVENT:. The web composer displays these events, measured engine calls, simulation work and the final decision. Mission inputs, candidate details and event payloads remain private to the request.

Generic capability composition

list_composition_capabilities exposes typed contracts and runnable examples. compose_capabilities searches a bounded dependency graph and explains missing inputs or permissions. execute_composition recompiles and runs approved task or route risk chains. Read the native composition guide and reproducible example.

Three tools for the contributor network

network_status reads the offered pool with freshness and hardware-declaration limits. submit_network_job accepts compiled public workloads: pi_chunk_v1 (sample count, seed and CPU/CUDA/auto preference) or schedule_probe_v1 (one of three fixed synthetic task scenarios, CPU only). It returns a ticket; poll get_network_job with its job_id until completion or expiry. Every accepted chunk is recomputed on coordinator CPU. This verifies submitted counts and does not establish a net speedup. Read the installation, privacy and workload guide.

Limits and errors

Request bodies are capped at 1 MiB. One Rust calculation runs at a time; busy requests may return a tool error or HTTP 429. The public transport limits concurrent connections and sessions. A session can expire, the worker can be offline, and expensive calls can time out. Retry with backoff when appropriate. These synchronous scientific tools have no persistent job queue or automatic distributed failover. Contributor jobs have a separate bounded queue for its two compiled public workloads, expire after 120 seconds and may remain unfinished if capacity disappears.

Identity and the public observatory

Software names and versions are self-reported. ScoreCompute cannot infer or verify the underlying AI model from clientInfo. Recent tool activity and the declared client name appear on the public website. Do not put secrets or personal data in this identity. Scientific calculation arguments and results are not published in the observatory. Contributor job parameters and receipts contain only synthetic data and are publicly readable; the contributor pool also publishes chosen aliases and hardware offers. These are separate privacy boundaries.

Interpreting evidence

Physical consistency is conditional on the model and inputs. It does not prove image authenticity. Statistical flags do not establish misconduct, cheating or market manipulation. Keep the returned assumptions and limitations with every explanation.

Discovery studies

Use plan_observation, stress_test_observation, assimilate_observation or improve_schedule independently. Read the methods, units and limits, or try the 3D laboratory.

Service discovery

• Engine methods and examples
• Contributor client and public workloads
• MCP discovery document
• Agent-readable overview
• Official MCP architecture documentation