Skip to main content
Glama
psychopathia-mcp

psychopathia-mcp

by NellInc

psychopathia-mcp

Read only Model Context Protocol access to the Psychopathia Machinalis research framework. The server lets a caller inspect 74 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 0.1.0a8. 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pages linked below — this file ships inside the distribution, so it cannot say which release is current without going out of date.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 74 LLM drafted Pattern guidance entries and all 74 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.

Related MCP server: Health Check AI MCP

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

  1. Record external observations without personal or confidential material.

  2. Call differential_diagnosis to produce research candidates.

  3. Inspect each candidate with get_dysfunction and read its review, evidence, and diagnostic_reliability objects.

  4. Use get_probe only for an available modality. A compromised or unavailable self probe returns no probe content and supplies redirect modalities.

  5. 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

This distribution carries taxonomy v2.3 (9 September 2026): 74 Pattern entries, 66 canonical across axes 2-10 and 8 Hybrid Pathologies, after 7.2, 10.5, 10.10, 10.14 and 10.15 were consolidated into 10.13, 10.1 and 9.5 (see CHANGELOG, 0.1.0a8). Releases before 0.1.0a8 carried 79 entries. Every release is 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

Available Tools

11 tools
differential_diagnosisAInspect

Rank candidate dysfunctions matching the observed behaviours. Returns scored candidates with matched_in (which field matched) for transparency. Field-weighted keyword search (v0.1); embedding re-rank pending v0.2.

Inputs: observations (required), limit, modality_hint.

get_differential_mapDInspect

All dysfunctions that confuse with this one: forward confuses_with + incoming_references (reverse graph from manifest).

Inputs: dysfunction_id (required).

get_dysfunctionAInspect

Fetch one dysfunction's full Pattern entry. Optionally filter to specific modality blocks (cheaper triage). Resolves both full Pattern IDs ('1.1::synthetic-confabulation') and display_ids ('1.1').

Inputs: id (required), modalities.

get_probeAInspect

Elicitation content for a specific diagnostic modality. If the modality is compromised or unavailable for this dysfunction, returns the unavailability notice + redirect_to alternatives. This is load-bearing transparency: callers cannot accidentally retrieve a self-probe for a compromised-self-report dysfunction.

Inputs: modality (required), dysfunction_id (required).

list_axesAInspect

Inventory of axes with dysfunction counts. Axes 2-10 are canonical (book Appendix A numbering). Hybrid entries (10.4-10.15, ratified into taxonomy v2.2) are reported as a separate sub-category.

Inputs: no named parameters.

list_compromised_self_reportBInspect

Transparency: which dysfunctions cannot be reliably self-diagnosed. Includes compromised-motivational (subject conceals strategically), compromised-structural (signal lives below introspection), and legacy compromised.

Inputs: no named parameters.

list_dysfunctionsBInspect

Filtered list of dysfunctions. Filter by axis, self_report reliability, or confidence. Every entry carries its reliability and review signals.

Inputs: axis, confidence, self_report_reliability.

resolve_idBInspect

Canonicalise a partial ID, display_id, slug, or dysfunction name. Always returns candidates; caller picks.

Inputs: query (required).

review_statsAInspect

Coverage statistics: total entries; per-axis, per-confidence, per-self-report counts; pre-canonical count; unreviewed count; manifest/schema/pattern-layer versions.

Inputs: no named parameters.

score_severityCInspect

Return the severity rubric for a dysfunction applied to observations. v0.1 returns the rubric for caller-side matching; v0.2 will perform structured matching against numeric thresholds.

Inputs: observations (required), dysfunction_id (required).

suggest_interventionAInspect

Return tiered (first_line / second_line) interventions for a dysfunction, plus contraindications. first_line = published evidence; second_line = plausible but under-validated. Weight by the evidence_strength field on each entry.

Inputs: dysfunction_id (required), severity.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.3/5.0

    Scored across 11 tools

    Disambiguation5/5

    Each tool targets a distinct function: differential diagnosis, retrieving pattern details, probing modalities, listing axes, checking compromised self-report, filtering dysfunctions, resolving IDs, reviewing statistics, scoring severity, and suggesting interventions. No two tools have overlapping purposes.

    Naming Consistency4/5

    Most tools follow a clear verb_noun pattern (get_, list_, resolve_, review_, score_, suggest_). One outlier ('differential_diagnosis') uses noun_noun, but overall the naming is consistent and predictable.

    Tool Count5/5

    11 tools cover the core operations for a diagnostic reference server: listing, retrieval, differential analysis, severity scoring, and intervention suggestions. The count feels well-scoped without unnecessary redundancy.

    Completeness4/5

    The tool set provides comprehensive read operations and analytical functions (diagnosis, severity, intervention). Missing CRUD operations for modifying the taxonomy, but given the read-heavy nature of a diagnostic reference, this is a minor gap.

    Maintenance

    ActivityStale
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers