📚 Agentic Paper

Multi-agent peer review · structured outputs · per-call audit log · v{{ version }}

⚠️

Auto-simulation mode — self-review only.

This tool runs LLM agents that simulate a peer-review committee. Use it on your own manuscripts to dry-run feedback before submitting to a real journal. It is not a substitute for human peer review, and should not be used to assess another author's paper without their explicit consent. LLM agents can hallucinate, miss context, or carry training biases — treat every recommendation as a starting point for human judgement, never the final word.

📄

Drop a PDF or text file

…or click to browse.

🔁 Compare against an earlier version (optional)

Drop the previous draft (v1) here. Agentic_Paper will diff it against the file above (v2), and the Revision Assessor agent will judge whether the revision is substantive, cosmetic, or a regression. Leave blank for a single-version review.

🔑 Bring your own API keys (optional)

Keys are sent with this upload, held in RAM only for the duration of this run, and never written to disk or logs. Leave a field blank to use the key the server has configured. Use HTTPS in any non-local deployment — these inputs are type=password but the browser still transmits them over the connection.

⚡ Quality vs cost preset (optional)

Pick a tiered model bundle. Each preset routes the 12 agents across high / standard / basic complexity tiers with different model choices. Cost estimates below assume ~50k input + 5k output tokens per agent call (a typical 20-30 page paper).

🤖 Pick a primary model for this run (optional)

By default, the routing in config.yaml picks a different model per agent based on complexity. Pin a single (provider, model) here to force every agent to use it for this run (tier_high keeps adaptive thinking on capable models).

Once you start, the orchestrator runs the 12 reviewer agents in parallel, then coordinator → summary → editor. The next page shows live activity from the audit log; the markdown report appears under Show report when the run completes.

With no BYOK keys, the run uses whichever providers the server has configured (see config.yaml). With BYOK, your keys override the matching provider for this single run.