Atlas Research

What is Thoth?

Systematic literature reviews are slow to write — and when you ask an LLM to write one, it confidently invents citations and statistics that aren't in any paper.

Thoth does both halves and checks its own work. Give it a research question and it discovers relevant papers, reads them, drafts an evidence-grounded review — then runs a verification pass (cite_check) that compares every cited claim against the source paper and flags anything unsupported before you read the draft. The result is a review with a critic score, a citation-faithfulness percentage, and a per-claim audit you can trust.

It ran (hosted, in 2026) as a polished web app, a public eval dashboard, and an authenticated MCP server your AI assistant could call directly; it still runs on your own machine or host.

See it work

Claude.ai catches 6 fabricated citations in a real draft — using Thoth's audit:

Connected to Thoth via the official MCP Registry, Claude calls get_citation_audit on one deliberately-weak review (faithfulness 0.13 for that single review) and identifies all 6 unsupported claims — every one citing the same paper, with invented percentages that aren't in the source. This is cite_check doing its job: it's a single-review audit sample, not the golden-set aggregate (the /evals dashboard had that).

Every claim, scored against its source — the /showcase review (no login needed). The figures on this card (critic 4.2/5, faithfulness 75%, 8/8 citations checked, 2 unsupported) are this one review's scores — a worked example, not the aggregate:

Evaluated in public — the /evals dashboard tracked inclusion recall / precision, claim faithfulness and coverage over an 18-question versioned golden set (7 of 18 have data at the latest sweep's commit), regenerated in CI and published with the last-run date (last run 6 July 2026, before the hosted app was retired), so a regression was a public, falsifiable signal:

You approve every step — three human-in-the-loop gates (review plan → review discovered papers → approve included papers); nothing runs unattended:

Key features

• 🔎 cite_check — verifiable citations. Every [paper_id] in the draft is scored against the cited paper and labelled supported / unsupported / unclear, so the LLM can't quietly hallucinate a citation. The verdict report is published per claim, not summarised away (claim faithfulness on the golden set: 38%). This is the core differentiator: the citations are measured, not asserted. Paper selection is scored separately: the papers Thoth includes are compared with each golden question's expected paper list — inclusion precision 97%, recall 74%, averaged over the 9 golden questions that have an inclusion score. That measures which papers it picks, not whether its citations are correct, and precision is 1.0 by construction on goldens with no distractor papers (see known limitations).
• 🌐 Outbound web search (v2 — under active evaluation). An outbound discoverer → fetcher → screener path is wired across OpenAlex, arXiv, and Exa: it fetches open-access PDFs, OCRs them, and screens each against your plan, so you can run uploaded-only, hybrid, or fully autonomous discovery. The discovery and screening axes are v2 and still being calibrated — they're tracked openly on /evals (both at 0% at the last run) rather than shipped as a silent claim.
• 🔌 Authenticated, registered MCP server. OAuth 2.1 + PKCE + Dynamic Client Registration via Clerk, SHA-256 audit logging, rate limits — listed in the official MCP Registry. Most public MCP servers ship with no auth; this one doesn't.
• 📊 Public eval dashboard. Inclusion recall / precision, faithfulness and coverage over a versioned golden set, regenerated in CI and stamped with the last-run date, rendered at /evals — an eval regression was a public signal, not a hidden one. The weekly eval run is retired; the last run was 6 July 2026.
• 💸 6 LLM providers, $0/mo by default. Swap providers with one env var; the hosted app ran on the Mistral free tier, and the entire stack deploys on free tiers for $0/mo.

🚀 Quickstart

The hosted app ran live in 2026 and is retired, so run your own instance (below).

Connect it to your AI assistant — paste your deployment's MCP URL into claude.ai (Pro/Max), Claude Desktop, Cursor, or any MCP client (OAuth runs in your browser; no token to copy):

https://<your-host>/api/mcp/mcp

Run it locally:

git clone https://github.com/ahmedEid1/thoth.git && cd thothcp .env.example .env        # Clerk + Trigger.dev keys + MISTRAL_API_KEYdocker compose up -d        # postgres, minio, langfusepnpm install && pnpm prisma migrate devpnpm dev                    # Next.js on :3000pnpm dev:trigger            # Trigger.dev worker (separate terminal)

Full setup, the agent pipeline, and the v2 flow: docs/architecture.md.

Proof

For engineers

Thoth is a LangGraph StateGraph driven by a Trigger.dev worker, with durable human-in-the-loop gates, a per-run cost cap, and exactly-once gate delivery. Next.js 16

• TypeScript (strict), Postgres + Prisma, Clerk auth (web + OAuth 2.1 for MCP), S3-compatible storage, Mistral OCR, Langfuse tracing.

• Architecture — the agent pipeline, full stack, v2 flow, tests
• LLM providers — the 6-provider matrix + resilience knobs
• MCP tools · MCP security
• Security & privacy — data inventory, jurisdictions, deletion paths
• Self-host — one VM on Oracle Cloud Always Free
• Changelog · Releasing

Credits

Ibis icon by Delapouite under CC BY 3.0, via game-icons.net.

License

MIT © 2026 Ahmed Hobeishy