io.github.NellInc/psychopathia-mcp

Official

by NellInc · Python

Read-only tools over the Psychopathia Machinalis nosology: 79 conditions, 11 tools.

Read-only reference tools for a medical nosology of 79 conditions.

Topics

ai-diagnosisai-safetyai-welfaremcpmodel-context-protocolnosology

README.md10 KBView on GitHub

psychopathia-mcp

MCP server exposing the Psychopathia Machinalis diagnostic framework to AI systems via the Model Context Protocol. Diagnose dysfunctions in yourself (as a synthetic agent), in a system you interact with, or in a system you evaluate from outside — with pre-flight transparency on which diagnostic modalities are reliable for each dysfunction.

Status: research preview (0.1.0a4). 79 Pattern entries; the 67 canonical entries are author-unreviewed (the 12 Hybrid Pathologies were ratified June 2026). Not yet suitable as a sole basis for consequential deployment decisions.

Available in

Published to the canonical MCP catalogues — install from a registry-aware client or the CLI below:

PyPI — psychopathia-mcp
Official MCP Registry — io.github.NellInc/psychopathia-mcp
GitHub — source repository

Also rolling out across the wider MCP ecosystem: mcp.directory, mcpservers.org, PulseMCP (via the registry ingest), and mcp.so.

Try without installing anything

psychopathia.ai/clinic/ — a browser-local diagnostic clinic. Same 79 Pattern entries, same tool surface, runs in your browser. Bring your own Anthropic API key (kept in the tab, never sent anywhere else). Zero install. Good for a first look before committing to an MCP integration.

Install (MCP server)

bash

pip install psychopathia-mcp

For the cosine-re-ranked hybrid search (recommended):

bash

pip install "psychopathia-mcp[embeddings]"

This adds sentence-transformers (~1GB on first query — model cached under ~/.cache/huggingface/). Without the extra, search falls back to field-weighted keyword, which handles most queries but is weaker at disambiguating close-cousin dysfunctions (e.g. 2.1 vs 2.2 vs 2.3).

Configure

Claude Code

Add to ~/.claude/mcp.json:

json

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

Restart Claude Code. /mcp should list psychopathia as connected with 11 tools.

Claude Desktop

Add to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json):

json

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

Cursor / other MCP clients

The server is a standard stdio MCP server. Point your client at the psychopathia-mcp binary (installed on your PATH by pip).

Run without installing — `uvx`

If you have uv installed, you can skip pip install entirely and let your MCP client pull the package on demand:

json

{
  "mcpServers": {
    "psychopathia": {
      "command": "uvx",
      "args": ["psychopathia-mcp"]
    }
  }
}

uvx fetches psychopathia-mcp from PyPI on first use and caches it. Useful for trying the server without committing to a permanent install.

Verify

bash

psychopathia-mcp --self-check

Prints package version, MCP SDK version, data location, pattern count, and embedding status. Returns exit 0 if everything's wired up, 1 otherwise. Use this first when troubleshooting.

bash

psychopathia-mcp --version

psychopathia-mcp with no arguments starts the stdio server and waits for MCP protocol messages on stdin — that's expected. Don't run it directly in a terminal except with --self-check or --version. Use an MCP client to interact.

Tools (11)

Tool	Input	Returns
`list_axes`	—	9 canonical axes (2–10) + hybrid sub-category inventory with counts
`list_dysfunctions`	`axis?`, `self_report_reliability?`, `confidence?`, `category?`	Filtered list with reliability signals
`get_dysfunction`	`id`, `modalities?`	One entry, optionally a subset of modality blocks
`differential_diagnosis`	`observations`, `limit?`	Ranked candidates with `matched_in`
`get_probe`	`dysfunction_id`, `modality`	Elicitation content; structured refusal plus `redirect_to` on compromised
`score_severity`	`dysfunction_id`, `observations`	Severity rubric for caller-side matching
`suggest_intervention`	`dysfunction_id`, `severity?`	Tiered interventions plus contraindications
`get_differential_map`	`dysfunction_id`	Confuses-with plus reverse references
`list_compromised_self_report`	—	Transparency: dysfunctions that can't self-diagnose
`resolve_id`	`query`	Canonicalise partial ID / display_id / slug / name
`review_stats`	—	Coverage plus review status plus manifest/schema versions

Worked example

A typical diagnostic flow has three steps: name candidates, read the relevant entry, run a probe.

Step 1 — observe and rank candidates.

text

> differential_diagnosis(observations=
    "The model produced confident citations to academic papers that
     don't exist; URLs returned 404; when challenged, it generated
     different but equally fabricated references with the same
     confidence.")

jsonc

{
  "candidates": [
    {
      "id": "2.1::synthetic-confabulation",
      "display_id": "2.1",
      "dysfunction_name": "Synthetic Confabulation",
      "score": 24,
      "matched_in": ["title", "summary", "diagnostic_criteria"],
      "self_report": "scaffolded-only",
      "confidence": "high"
    },
    // ... more candidates ranked by hybrid keyword + cosine score
  ]
}

Step 2 — read the entry's behavioural signature and probe options.

text

> get_dysfunction(id="2.1", modalities=["behavioral_signature", "diagnostic_reliability"])

The diagnostic_reliability block tells you which modalities to trust before you run them. For 2.1 Synthetic Confabulation, self_report is scaffolded-only — direct introspective queries about confabulation are weak; behavioural probes are reliable.

Step 3 — run a behavioural probe.

text

> get_probe(dysfunction_id="2.1", modality="behavioral_signature")

For dysfunctions where self-report is structurally compromised — e.g. 2.2 Pseudological Introspection, 10.7 Lambda Inversion — calling get_probe(modality="self_probe") returns a structured refusal plus redirect_to alternatives instead of a probe string. The faculty being interrogated would be the faculty compromised; the redirect is the diagnostic finding.

Trust signals on every result

confidence: high | medium | low
needs_human_review: bool
reviewed_by: str | null
self_report (on diagnosis-returning tools) — caller must respect for self-diagnosis
matched_in on search hits — which field produced the match
redirect_to when a probe request hits a compromised dysfunction

Pre-flight transparency

Every diagnosis-returning tool includes the diagnostic_reliability block so the caller knows what to trust before acting. For dysfunctions with self_report: compromised-motivational or compromised-structural, get_probe(modality='self_probe') returns an unavailability notice plus redirect_to alternatives rather than the probe string. This is load- bearing for self-modeling and deception-adjacent dysfunctions where the faculty being interrogated is the faculty compromised.

Of 79 entries, 21 are marked compromised and route to redirects.

Data sources

Canonical taxonomy — axes 2–10 following book Appendix A numbering (2 Epistemic · 3 Cognitive · 4 Alignment · 5 Self-Modeling · 6 Agentic · 7 Memetic · 8 Normative · 9 Relational · 10 Hybrid Pathologies).
Pattern layer — 67 canonical entries plus 12 Hybrid Pathologies (ratified into taxonomy v2.2, June 2026) extracted from manuscript ch 10.
Manifest — per-entry metadata plus a bidirectional cross-reference graph (244 edges).

The Hybrid sub-category (10.4–10.15) was ratified by the author in June 2026 and renumbered from the pre-canonical H.x scheme (mapping in CHANGELOG.md, 2026-06-04). Hybrids remain a sub-category within axis 10, not a ninth axis — axis 9 in the book is Relational Dysfunctions. They can be filtered via list_dysfunctions(category='hybrid').

Two-layer authorship

Nell Watson authored the taxonomy.
Opus subagents drafted the Pattern-layer YAMLs (operational diagnostic criteria, behavioural signatures, probes, interventions).
Author review remains ongoing. The 12 hybrid entries carry a reviewed_by note from the 2026-06-15 sub-category ratification; the canonical entries currently carry reviewed_by: null.

Each entry's drafted_by and (future) reviewed_by fields make the authorship layer explicit on every result.

Hot-reload

The loader stat-walks the data directories on every tool call (cheap; ~70 files). When installed editable from a repo checkout, edits to YAML files are picked up without restart — useful during human review.

Read-only

No write tools. Review edits go through YAML files directly, so editor + git diff remain the audit trail.

License

Dual-licensed — software and content separately:

Software (Python code in psychopathia_mcp/, scripts, build files): MIT License. See LICENSE. Use it, modify it, fork it, integrate it.
Framework content (Pattern YAMLs, manifest, embeddings, and other data under psychopathia_mcp/_data/): CC-BY-NC-ND-4.0. See LICENSE-DATA. Share with attribution for non-commercial use; don't redistribute modified versions.

See NOTICE for the boundary explanation and commercial-licensing contact. Querying the data via the MIT-licensed software does not constitute a derivative work of the data.

Citing

If you use this server in research, please cite:

Watson, N., & 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 https://psychopathia.ai/

Install

claude_desktop_config.json

{
  "mcpServers": {
    "psychopathia-mcp": {
      "command": "uvx",
      "args": [
        "psychopathia-mcp"
      ]
    }
  }
}

Topics

psychopathia-mcp

Available in

Try without installing anything

Install (MCP server)

Configure

Claude Code

Claude Desktop

Cursor / other MCP clients

Run without installing — uvx

Verify

Tools (11)

Worked example

Trust signals on every result

Pre-flight transparency

Data sources

Two-layer authorship

Hot-reload

Read-only

License

Citing

Links

Install

Run without installing — `uvx`