Aurelius

A fact-checked research MCP server. Aurelius gives any MCP-capable app — Claude (Desktop / Code / claude.ai), Gemini CLI, Cursor, and (via a remote deployment) ChatGPT — a set of research tools that verify every citation against real scholarly databases (OpenAlex, Crossref — DOI-backed and retraction-aware) and every claim against live web sources before presenting it. No more hallucinated papers, no more silently-cited retracted studies.

Aurelius grew out of a multi-agent research framework and distills its best idea into a portable tool server: screen a topic → draft → fact-check → revise.

Writing the paper yourself, in LaTeX? Aurelius-IDE is the sibling project: a language server that runs these same retraction-aware checks inline in your editor, as you type, rather than through a chat client. It uses this package as its optional verification backend. Same checks, different surface.

Why this design solves the "API cost" problem

By default Aurelius runs in host-driven mode: the app you connect it to (Claude, Gemini, etc.) uses its own model to reason and write, and Aurelius just supplies the research and fact-checking tools. That means Aurelius needs no LLM API key of its own — the tokens are covered by your existing Claude/Gemini/ChatGPT subscription. Citation verification runs against OpenAlex and Crossref — both free, both keyless. The only optional key is Tavily, used for general web_search and as a fallback when a citation isn't indexed in either scholarly database (free tier available).

There's also an optional autonomous mode (autonomous_research / aurelius-research) where Aurelius drives its own LLM — that one needs an LLM API key with quota.

Related MCP server: truth-anchor-agent

Install

pip install aurelius-mcp

The bare name aurelius was already taken on PyPI, so the package ships as aurelius-mcp. The import name (import aurelius) and the CLI command (aurelius) are unchanged.

This provides two commands:

• aurelius — launch the MCP server (stdio). This is what MCP clients run.

• aurelius-research "<topic>" — run one autonomous research job from the terminal.

If aurelius isn't found (the pip scripts dir may not be on your PATH — common on Windows), use the equivalent module form anywhere a command is expected: "command": "python", "args": ["-m", "aurelius"].

Get a Tavily key (optional — for general web search)

Citation verification (verify_citation, verify_claims) needs no key — it runs against the free, keyless OpenAlex and Crossref APIs. A Tavily key is only needed for web_search (general factual-claim evidence) and as a fallback when a citation isn't indexed in either scholarly database. Create a free key at https://tavily.com and expose it as TAVILY_API_KEY (see the config snippets below, which inject it into the server's environment).

Connect it to your app (local / stdio)

Claude Desktop

Edit claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "aurelius": {
      "command": "aurelius",
      "env": { "TAVILY_API_KEY": "tvly-your-key" }
    }
  }
}

Restart Claude Desktop. See examples/claude_desktop_config.json.

Claude Code

claude mcp add aurelius --env TAVILY_API_KEY=tvly-your-key -- aurelius

Cursor

Add to ~/.cursor/mcp.json (or the project .cursor/mcp.json):

{
  "mcpServers": {
    "aurelius": { "command": "aurelius", "env": { "TAVILY_API_KEY": "tvly-your-key" } }
  }
}

Gemini CLI

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "aurelius": { "command": "aurelius", "env": { "TAVILY_API_KEY": "tvly-your-key" } }
  }
}

Then just ask: "Use Aurelius to research the historical correlation between GDP growth and unemployment, and verify every citation."

Seeing it catch a bad citation

A real run on "the historical correlation between GDP growth and unemployment (Okun's law)": Claude drafted the paper, then called verify_citation on every reference.

Nothing unverifiable made it into the final draft. That's the whole point.

Seeing it catch a retracted paper

verify_citation doesn't just check that a paper exists — it checks OpenAlex's live retraction registry. A real call against the (in)famous Wakefield MMR-autism paper:

verify_citation("Wakefield, A. J. et al. (1998). Ileal-lymphoid-nodular hyperplasia, "
                 "non-specific colitis, and pervasive developmental disorder in children.")

{
  "verdict": "retracted",
  "is_retracted": true,
  "confidence": "high",
  "source": "openalex",
  "matched_work": {
    "title": "RETRACTED: Ileal-lymphoid-nodular hyperplasia, non-specific colitis, ...",
    "doi": "10.1016/s0140-6736(97)11096-0",
    "year": 1998
  },
  "notes": "Retracted work — flagged by openalex. Do not cite."
}

is_retracted is always a top-level field — impossible for a host model to miss or rationalize past. A scholarly index can return several records for the same paper (the original, a retraction notice, clean-looking duplicates); Aurelius specifically resolves ties in favor of surfacing the retraction rather than picking whichever record looks cleanest.

Seeing it catch a mis-attributed citation

A title match alone is not a verification. Aurelius corroborates the cited author and year against the matched record, so it catches the subtle case a title-only checker waves through:

verify_citation("Okun, A. M. (1962). Potential GNP: Its Measurement and Significance.")

{
  "verdict": "unverified",
  "author_match": false,
  "match_score": 1.0,
  "matched_work": { "authors": ["Charles I. Plosser", "G. William Schwert"], "year": 1979 },
  "notes": "Found a work with this title but different authors (found: Plosser, Schwert; cited: Okun) — likely not the paper you cited."
}

The title matches perfectly (1.00), but the only indexed record with that title is a 1979 paper by Plosser & Schwert — not Okun's 1962 original. A title-only checker reports ✓; Aurelius reports the truth and hands back the corrected_citation for the record it actually found. When a citation carries a DOI or arXiv id, it's looked up directly for an exact match.

Tools

Outputs are written to ~/aurelius_output/ in your home directory (override with AURELIUS_OUTPUT_DIR) — never to the process's current working directory, since MCP clients often launch the server from a location you can't write to.

Long-form papers (20–80+ pages)

Call plan_paper_length(target_pages=40) for a section-by-section word-count budget, then draft and verify_claims one section at a time, appending each with save_draft(content, filename, append=True) so the host model never has to resend the whole accumulated document. See SKILL.md for the full workflow.

A note on polish_prose

It's a readability pass on already-verified content — it fixes stiff, repetitive LLM phrasing (hedging chains, transition-word stacking, tricolon padding) while preserving every citation, number, and claim verbatim. It is explicitly not an AI-detector-evasion tool; pairing that with long-form academic paper generation would enable academic dishonesty, which is out of scope for a project whose entire premise is showing verifiable receipts.

The Claude skill

skill/aurelius/SKILL.md teaches a host model the exact screen → plan → draft → verify → polish → save workflow, including the long-form (section-by-section) path. Drop it into your Claude Code/Agent skills so the model uses the tools rigorously.

Autonomous mode (optional, needs an LLM key)

export OPENAI_API_KEY=sk-...          # or ANTHROPIC_API_KEY / GOOGLE_API_KEY
export TAVILY_API_KEY=tvly-...
aurelius-research "Health effects of microplastics in drinking water" --model gpt-4o-mini-2024-07-18 --rounds 2

Provider is auto-detected from the model name (gpt-* → OpenAI, claude-* → Anthropic, gemini-* → Google).

Orchestration mode — the multi-stage agent DAG (--graph)

Beyond the linear loop, Aurelius can run a staged research DAG driven by a swarm of specialized agents (literature mining → a parallel hypothesis swarm → feasibility screening → experiment design/code → citation verification → adversarial review → drafting → LaTeX → proof-of-rigor). Every agent action is logged to an audit trail and each stage is checkpointed under ~/aurelius_output/sessions/, so a run is fully inspectable and resumable.

aurelius-research "Effect of sleep duration on reaction time" --graph

Or call the autonomous_research_graph MCP tool from any client. It's built in-house — no LangGraph/LangChain — reusing the same retraction-aware citation verification as the rest of Aurelius.

Code sandbox & p-hacking audit (Phase 2). The DAG statically audits the generated analysis code for questionable-research-practice signals (uncorrected multiple comparisons, missing random seed, post-hoc outlier removal, optional stopping, HARKing, selective reporting) and reports a risk score. Add --sandbox to also execute that code in a hardened, network-less Docker container (CPU/mem/pids caps, read-only fs, dropped capabilities, non-root, timeout) — opt-in because the code is model-written, and a graceful skip if Docker isn't present:

aurelius-research "your topic" --graph --sandbox

Cryptographic Proof-of-Rigor (Phase 3). Each run emits a signed, tamper-evident proof bundle: a SHA-256 content hash of the evidence ledger + full audit trail, signed with ed25519 (or HMAC if you set AURELIUS_PROOF_HMAC_SECRET), written to ~/aurelius_output/proofs/ and independently checkable with aurelius.proof.verify_proof(...). Optional IPFS pinning (set PINATA_JWT) and optional on-chain anchoring (pip install aurelius-mcp[chain] + set AURELIUS_CHAIN_RPC / AURELIUS_CHAIN_PRIVATE_KEY) layer on top; both are graceful no-ops when unconfigured.

Publishing, DeSci & memory (Phases 4–6). The DAG also packages a submission-ready preprint bundle (compile-ready LaTeX, a DOI-backed references.bib built from the Evidence Ledger, and a per-server checklist for arXiv/bioRxiv/medRxiv) — submission itself stays manual by design, since those servers require human endorsement/moderation. It runs a patent-freedom screen (PatentsView, a screening aid — not legal advice), and it has episodic memory: every run is recorded with derived lessons and relevant past runs are recalled at the start of new ones. check_compliance remains an honest placeholder — see ARCHITECTURE.md for the full status.

Platform support (honest status)

License

MIT