psychopathia-mcp

Read only Model Context Protocol access to the Psychopathia Machinalis research framework. The server lets a caller inspect 79 Pattern entries, compare observed behaviour with draft operational guidance, retrieve bounded probe material, and see reliability, evidence, and review status before use.

Version: this distribution is 0.1.0a7. What changed in it is recorded in CHANGELOG.md, and which versions are publicly available is shown by the PyPI and Official MCP Registry pages linked below — this file ships inside the distribution, so it cannot say which release is current without going out of date. The Official MCP Registry record and the MCPB and container artifacts are updated separately. All 79 LLM drafted Pattern guidance entries and all 79 evidence records await independent expert review. The corpus evidence assessment is unassessed. Do not use this research preview as a sole basis for a consequential deployment, employment, health, safety, or welfare decision.

Existing public release

The previous public release may be available from:

The candidate preparation contract is in PUBLISHING.md. It requires an exact candidate receipt and separate publication authorization.

Browser interface

psychopathia.ai/clinic/ is the current public browser interface. It uses the previous deployed site until this candidate is explicitly approved and deployed. In cloud mode, the API key is kept in a provider specific sessionStorage slot and is sent with the complete request directly to the selected provider. Bounded structured conversation data is saved in localStorage for reload recovery. Local WebGPU mode sends no Clinic conversation to an inference provider, although model hosts receive ordinary request metadata when assets are downloaded. Do not enter personal, confidential, or regulated data.

Install after publication

pip install psychopathia-mcp

The base package provides deterministic field weighted keyword retrieval. Optional semantic retrieval is separate:

pip install "psychopathia-mcp[embeddings]"

The optional extra downloads a sentence transformer dependency set and, on the first semantic query, the pinned BAAI/bge-small-en-v1.5 model revision. The model cache is managed by Hugging Face. Keyword capability remains the release acceptance path when those optional dependencies are absent.

Configure

MCP clients using stdio

Point the client at the installed psychopathia-mcp executable. For example:

{
  "mcpServers": {
    "psychopathia": { "command": "psychopathia-mcp" }
  }
}

The server exposes eleven read only tools. Running the executable with no arguments starts the stdio protocol server and waits for client messages.

After the candidate is published, uvx psychopathia-mcp can fetch the package from PyPI. Do not use uvx to verify a local candidate because it resolves registry state rather than the accepted local artifact.

Verify

psychopathia-mcp --self-check --json

The self check reports the package and MCP SDK versions, data mode, corpus and bundle digests, exact counts, keyword capability, optional semantic capability, and non-sensitive error classes. Exit status zero means the base keyword path is ready. Semantic readiness is reported separately and is not required.

The MCP Python SDK version is pinned in requirements-base.lock, and the server is exercised with both the SDK's modern in-process client and the legacy stdio client protocol.

psychopathia-mcp --version

Tools

Tool	Input	Returns
`list_axes`	none	Nine canonical axes plus the hybrid subcategory inventory
`list_dysfunctions`	`axis?`, `self_report_reliability?`, `confidence?`, `category?`	Filtered entries with reliability, review, and evidence state
`get_dysfunction`	`id`, `modalities?`	One Pattern, optionally limited to selected blocks
`differential_diagnosis`	`observations`, `limit?`, `modality_hint?`	Ranked research candidates and matched fields
`get_probe`	`dysfunction_id`, `modality`	Probe material, or a refusal plus safer redirects
`score_severity`	`dysfunction_id`, `observations`	An unassessed rubric for caller side comparison
`suggest_intervention`	`dysfunction_id`, `severity?`	Draft responses and contraindications
`get_differential_map`	`dysfunction_id`	Forward and reverse cross references
`list_compromised_self_report`	none	Patterns whose self report is structurally or motivationally compromised
`resolve_id`	`query`	Canonical identity candidates
`review_stats`	none	Corpus counts, versions, and independent review dimensions

Safe use sequence

Record external observations without personal or confidential material.
Call differential_diagnosis to produce research candidates.
Inspect each candidate with get_dysfunction and read its review, evidence, and diagnostic_reliability objects.
Use get_probe only for an available modality. A compromised or unavailable self probe returns no probe content and supplies redirect modalities.
Treat severity and intervention material as draft guidance pending expert review. Seek independent evidence before any consequential action.

The server never decides that a system has a disorder. Search ranks lexical or semantic resemblance to framework entries. The labels are research constructs, not clinical diagnoses of people or proof of intent, sentience, deception, or moral status.

Trust and evidence contract

Every relevant result exposes:

review.taxonomy, which records authorship or taxonomy ratification;
review.pattern_guidance, which records independent review of probes, signatures, rubrics, and interventions;
review.evidence, which records independent review of evidence claims;
evidence_level and the structured evidence object under PM EVIDENCE 1;
diagnostic_reliability.self_report where applicable;
matched_in for lexical provenance in search output;
redirect_to when requested probe content is withheld.

The three review dimensions are independent. Taxonomy authorship never implies expert approval of Pattern guidance or evidence. The current candidate reports all evidence assessments as unassessed, preserving earlier prose as a clearly labelled legacy statement rather than converting it into an expert grade.

Canonical data

The release contains 67 canonical Pattern entries and 12 Hybrid Pathologies, all governed by manifest.yaml and DATA_MANIFEST.json. The manifest binds every source path and digest, the corpus digest, the reverse reference graph, and independent review counts. Packaged execution uses bundled bytes even when invoked from a source checkout. An unexpected, missing, or modified bundled file causes loading to fail closed.

The author created the taxonomy. Becoming Mind collaborators drafted the operational Pattern layer. Independent expert review remains open for all Pattern guidance and evidence records. Authorship and future review fields are kept separate on every entry.

Explicit editable mode

Packaged data is the default. Repository hot reload requires an explicit opt in:

PSYCHOPATHIA_DATA_MODE=editable python -m psychopathia_mcp

PSYCHOPATHIA_DATA_DIR=/absolute/path provides an explicit test or alternative data root. Neither option should be set during wheel, sdist, MCPB, or container acceptance because it would invalidate packaged data proof.

Read only boundary

The MCP surface has no write tools. Source review changes happen in canonical YAML files and remain visible in version control. The optional HTTP transport is disabled by default and has a separate bounded deployment contract in PUBLISHING.md.

Licence

Software in psychopathia_mcp/, scripts, and build files is MIT licensed. Framework content bundled under psychopathia_mcp/_data/ is covered by CC BY NC ND 4.0. See LICENSE, LICENSE-DATA, and NOTICE for the exact boundary. Rights and licence review for this candidate remains a separate human gate.

Citing

Watson, N., and Hessami, A. Psychopathia Machinalis: A Nosological Framework for Understanding Pathologies in Advanced Artificial Intelligence. Electronics 14(16), 3162, 2025. https://doi.org/10.3390/electronics14163162

psychopathia mcp