Documentation

Build with declarative graphs

Everything you need to compose, validate, and execute multi-step scientific computing & statistical analysis through the Euriklis MCP server or HTTP API.

Browse the docs

Start hereGetting startedInstall the MCP server, get an API key, and run your first DAG in under five minutes.ReferenceDAG specificationThe shape of a DAG request: inputs, ops, output, and the strict-reference rules.ReferenceOp catalogTensor primitives, graph algorithms, and optimization solvers — what each op accepts and returns.IntegrationMCP installPer-agent install instructions for Claude Desktop, Cursor, Codex, ChatGPT, and Zed.ReferenceHTTP APIUse Euriklis without an MCP host. POST a DAG, get back the result.ConceptsWhy DAGs?Why declarative graphs beat per-op tool calls when you compose multi-step scientific computing & statistical analysis.

DAG request shape

A DAG has three keys: inputs (named tensors or graphs supplied by the caller), ops (an ordered list of operations), and output (the label of the op whose result is returned).

Request

{
  "inputs": { "Z": "tensor[m,n]" },
  "ops": [
    { "op": "normalize",     "x": "Z",  "as": "Zn" },
    { "op": "scalarProduct", "x": "Zn", "k": 3.14159, "as": "out" }
  ],
  "output": "out"
}

Response

{
  "output": "out",
  "result": {
    "shape": [3, 3],
    "dtype": "f32",
    "data": "<base64>"
  },
  "stats": { "ops": 2, "ms": 4.2, "backend": "cpu-llvm" }
}

Validation rules

Strict referencesEvery x, scalar, or k field on an op must reference either a key in inputs or the as label of an earlier op. Undefined references throw — no auto-creation.

Pure and functionalAll ops are pure. There are no in-place mutations, no side-effects, and no void ops in the data plane. An op with no as label is a static error.

Static validationBefore execution, the validator runs shape inference, type checking, and cycle detection. Failures return a structured error and do not consume your call quota.

Why DAGs?

An agent that wants "normalize then multiply by π" can issue two MCP calls — one per op — but every round-trip is a latency hit and a chance for the agent to drift. Declaring the full computation as one DAG sends one request, validates the full plan up front, executes on a single backend, and returns the final tensor.

Each op stays a primitive (composable, testable) while the caller speaks at the level of intent. The compiler — LLVM for CPU, PTX for GPU — then fuses where it can.