psychopathia-mcp

MCP server exposing thePsychopathia Machinalisdiagnostic 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.

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), andmcp.so.

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.

pip install "psychopathia-mcp[embeddings]"

This addssentence-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).

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

Restart Claude Code./mcpshould listpsychopathiaas connected with 11 tools.

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

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

The server is a standard stdio MCP server. Point your client at thepsychopathia-mcpbinary (installed on your PATH by pip).

If you haveuvinstalled, you can skippip installentirely and let your MCP client pull the package on demand:

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

uvxfetchespsychopathia-mcpfrom PyPI on first use and caches it. Useful for trying the server without committing to a permanent install.

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.

psychopathia-mcpwith 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-checkor--version. Use an MCP client to interact.

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

> 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.")

{ "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.

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

Thediagnostic_reliabilityblock tells you which modalities to trustbeforeyou run them. For 2.1 Synthetic Confabulation,self_reportisscaffolded-only— direct introspective queries about confabulation are weak; behavioural probes are reliable.

> 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— callingget_probe(modality="self_probe")returns a structured refusal plusredirect_toalternatives instead of a probe string. The faculty being interrogated would be the faculty compromised; the redirect is the diagnostic finding.

- 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_inon search hits — which field produced the match
- redirect_towhen a probe request hits a compromised dysfunction

Every diagnosis-returning tool includes thediagnostic_reliabilityblock so the caller knows what to trust before acting. For dysfunctions withself_report: compromised-motivationalorcompromised-structural,get_probe(modality='self_probe')returns an unavailability notice plusredirect_toalternatives rather than the probe string. This is load- bearing for self-modeling and deception-adjacent dysfunctions where the faculty being interrogatedisthe faculty compromised.

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

- 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 inCHANGELOG.md, 2026-06-04). Hybrids remain asub-categorywithin axis 10, not a ninth axis — axis 9 in the book is Relational Dysfunctions. They can be filtered vialist_dysfunctions(category='hybrid').

- 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 areviewed_bynote from the 2026-06-15 sub-category ratification; the canonical entries currently carryreviewed_by: null.

Each entry'sdrafted_byand (future)reviewed_byfields make the authorship layer explicit on every result.

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.

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

Dual-licensed — software and content separately:

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

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

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/electronics14163162https://psychopathia.ai/

- Browser clinic (no install):https://psychopathia.ai/clinic/
- Documentation:https://psychopathia.ai/mcp.html
- Main project:https://psychopathia.ai/
- Issues / contact:https://psychopathia.ai/contact/

aTars MCP by aarna provides AI agents with structured access to crypto market signals, technical indicators, and sentiment analysis.