Honesty MCP

What honesty-mcp MCP server does

The honesty-mcp MCP server gives coding agents access to a collection of small TypeScript checks for evaluating claims made by AI products. The tools cover citation structure, evidence corroboration, payout and mutation invariance, trust scoring, anonymous-content authenticity, provenance language, claims freshness, audit-chain integrity, recommendation quality, and disagreement between an engine and a human.

The server is a thin adapter rather than a second implementation of these checks. Each tool defines an input schema and calls the corresponding export from an honesty kit. Results are returned as structured MCP responses that an agent can inspect or use in a development workflow.

Use the honesty-mcp MCP server to review generated answers, inspect supplied evidence, test ranking or decision behavior against scenarios, maintain an audit trail, or evaluate whether public claims remain supported by their evidence. It is not a general fact-checker: several checks are structural or heuristic and do not establish that an external source is true.

How it works

The server communicates through the stdio transport. It reads JSON-RPC requests from standard input and writes responses to standard output, while its startup message goes to standard error. It is intended for local MCP clients and does not expose an HTTP endpoint or listening port.

Inputs are supplied by the calling agent rather than loaded from a project automatically. For example, grounding checks receive draft text and source material, while invariance checks receive scenarios and a function to test. Static payout-import checking examines source text supplied to the tool and does not read files from disk.

Some results require careful interpretation. Grounding checks compare citation structure and lexical overlap, not factual truth. Corroboration reports evidence direction and graded results, but does not independently verify source independence. Invariance checks cover supplied scenarios and are not formal proofs across every possible input.

Setup and configuration

Run the honesty-mcp MCP server without a local install by using npx honesty-mcp, or install it in a project with npm install honesty-mcp. A Claude Code configuration can launch it with claude mcp add honesty-mcp -- npx -y honesty-mcp. Other clients using an mcpServers configuration can define the server with the npx command and -y honesty-mcp argument.

The server works with its standard tool set without environment variables. Three JavaScript-execution paths are disabled by default. Set HONESTY_MCP_ENABLE_CODE_EXEC=1 in the server environment to enable the runtime payout-invariance check, mutation-invariance check, and the optional live example behavior associated with scaffolding. Code execution should be enabled only after reviewing the inputs those tools will run.

Tools and capabilities

The honesty-mcp MCP server includes tools for:

• Checking sentence-level citation structure and identifying uncited, placeholder, or structurally invalid claims.
• Comparing evidence direction and grading corroboration across distinct sources.
• Testing payout sensitivity and mutation sensitivity with supplied scenarios.
• Scoring identified contributors and assessing whether anonymous sentiment appears organic.
• Finding certainty language without suitable provenance and classifying claims as current, stale, or unverified.
• Appending to and independently verifying a hash-chained audit log.
• Grading a recommendation against an observation log and measuring engine-versus-human divergence.
• Generating guidance and starter material for agent receipts and cost-governor patterns.

Several tools have explicit limits. Trust scoring shrinks sparse evidence toward a prior, while anonymous authenticity scoring uses heuristics and does not detect fabrication. Audit verification can detect different classes of mutation or truncation depending on the supplied options. Read each result together with its refusal codes, confidence, direction, or limitation fields where provided.

Limitations and notes

The checks are not universal guarantees. Lexical citation matching can reject faithful paraphrases or accept claims with substantial word overlap. A corroboration result does not prove independence, and a confirmed result may still describe confirmed contradiction if the direction is ignored. Scenario-based invariance testing cannot establish behavior for every possible configuration.

The server does not automatically browse sources, verify external facts, inspect files for static-import checks, or expose a web service. Agents must provide the material and scenarios to analyze. The underlying honesty kits are published npm dependencies, so installation resolves them from the public registry rather than local sibling repositories.