Overview

The platform addresses a class of problems in which the input is a sequence of well-typed mathematical operations and the desired output is the evaluation of that sequence on structured numerical data — matrices, tensors, tabular records, or discrete graphs. Representative use cases include dense linear algebra, iterative numerical procedures, statistical estimation, portfolio and financial analytics, and the integration of such computations into agent-driven workflows.

Rather than exposing individual functions as remote procedure calls, Euriklis accepts an entire computation as a single declarative artefact — a directed acyclic graph. This design minimises round-trip latency, admits static verification of the request prior to any execution, and permits the backend to schedule the constituent operations coherently on accelerated hardware.

Two equivalent invocation surfaces are exposed. Each accepts the same graph schema and produces results of identical form.

• HTTP API. A single JSON endpoint, POST /v1/compute, authenticated by an API key issued from the dashboard. Suited to integration with application code and back-office pipelines.
• Model Context Protocol (MCP) server. An out-of-process server distributed as an installable package. It exposes the same operational surface to language-model agents — Claude Desktop, Cursor, Codex, Zed, and any other MCP-aware host — permitting an agent to formulate and execute a complete computation without further prompting.

A request is a directed acyclic graph. Each node carries an ordered list of operations drawn from the platform's registry; each edge may carry a guard expression that determines whether traversal proceeds through it. Evaluation is initiated at a designated entry node and proceeds in breadth-first order along the set of edges whose guards are satisfied.

Prior to any execution, every submitted graph is subjected to static analysis: schema conformance, operation arity, dimensional and type inference, and cycle detection. Validation failures return a structured diagnostic and do not consume the caller's compute allowance. All operations are pure; there are no in-place mutations, no observable side effects, and no operation without a declared output label.

The registry is organised into thematic groups. A comprehensive listing, with signatures and complexity notes, is provided in the Operations catalog section.

• Dense linear algebra. Matrix multiplication, transposition, inversion; the standard decompositions — LU with partial pivoting, LDL, QR, SVD, and eigenproblems; Frobenius and induced norms; matrix balancing.
• Convolution and pooling. Two-dimensional convolution and correlation, pooling operators, and the customary activation functions (ReLU, LeakyReLU, ELU, softmax).
• Elementary and transcendental scalar functions. Trigonometric, hyperbolic, exponential, logarithmic, and inverse variants, together with the common statistical primitives.
• Physical and mathematical constants. A curated collection —π, e, Avogadro's number, the Boltzmann and Planck constants, the speed of light, and further quantities of interest.
• Financial-data operations. Market-data retrieval, portfolio aggregation primitives, and index-constituent queries.
• Data ingest. Tabular file handling for the injection of external data into the graph.

Consumption is measured in Euriklis Compute Units (ECU). Each executed operation incurs a fixed cost that is proportional to its arithmetic complexity and to the size of its operands. Failed validation does not deduct ECU. Current consumption, quotas, and per-key breakdowns are visible from the usage panel of the dashboard. Tariff details and plan comparisons are set out on the pricing page.

Newcomers to the platform are advised to proceed to the Quickstart, which walks through the issuance of an API key and the submission of a first computational graph. The DAG schema section provides the precise grammar of a request; the Operations catalog enumerates the operations available; the HTTP API reference documents endpoint semantics, authentication, error codes, and multipart file ingest; and the MCP integration section describes per-host installation for agent-driven use.