ComplyHat’s entire customer surface is a single MCP endpoint. You do not call a REST API, write a config file, or run a CLI — your MCP host (Claude Code, Codex, or any custom client) calls one of 14 compliance tools, and ComplyHat returns structured, audit-tagged citations. Your host brings its own reasoning; ComplyHat returns the deterministic numbers and document sections regulators want to see.Documentation Index
Fetch the complete documentation index at: https://docs.complyhat.ai/llms.txt
Use this file to discover all available pages before exploring further.
The request-response loop
Your host calls a tool
You ask your host agent to do something — “register this model,” “run a bias test,” “generate an EU AI Act report.” The host decides which tool to call and constructs the arguments. No ComplyHat-specific prompt engineering required on your part.
ComplyHat dispatches by entity and mode
Each tool name maps to an entity (e.g.
models, reports, bias_tests). Each call includes a mode that selects the exact operation — register, generate, run, and so on. This entity-and-mode pattern keeps the tool count at 14 regardless of how many underlying operations exist, which keeps your host’s tool selection accurate.Tests and document sections run deterministically
ComplyHat executes the relevant bias, drift, explainability, or adversarial tests — or renders the requested compliance document section — using only the data you supplied. There are no internal LLM calls. Every result is a function of your inputs, the engine version, and the random seeds recorded in the report.
Entity tools and mode dispatch
ComplyHat exposes 14 entity tools. Each entity groups a set of related operations as namedmode values. When your host calls reports with mode: "generate", it is invoking the report-generation operation. When it calls reports with mode: "get_pdf_url", it is requesting a download link for the generated PDF.
This design means your host sees a small, stable list of tool names. Host agents trained on large tool lists lose selection accuracy; keeping the surface to 14 tools keeps your host reliable.
A fifteenth tool — guidance — serves canonical skill and agent markdown on demand. Your host can call it to fetch the latest instructions for any ComplyHat auditor agent or operational skill without leaving its native interface.
Audit-tagged citations
Every prose field ComplyHat returns is prefixed with one of three tags:[EXTRACTED]
The claim is drawn directly from data or metadata you supplied. A reviewer can locate the source in the input.
[INFERRED]
The claim follows from the test results or document logic, but was not stated verbatim in your input. Reasonable inference, traceable to the evidence.
[AMBIGUOUS]
The input was incomplete or contradictory. A reviewer should inspect this field before the document enters an audit trail.
[AMBIGUOUS] field and resolve it with your subject-matter expert before submitting.
What ComplyHat does not do
ComplyHat runs zero internal LLM calls. It does not synthesize prose, paraphrase your inputs, or call your model’s prediction functionf. Your host agent owns all natural-language reasoning. ComplyHat owns the deterministic math — bias ratios, drift statistics, attribution scores — and the structured document templates.
Any finding in a ComplyHat report is re-derivable from the same inputs and engine version by a third party. ComplyHat persists the engine version, random seeds, and all input parameters alongside every report.
Next steps
Tool reference
All 14 entity tools, their modes, and example request and response payloads.
Quickstart
Add the MCP URL to your host and make your first tool call in minutes.