EurekaClaw: Multi-Agent AI Research Assistant for Theorem Proving and Paper Writing

What EurekaClaw does and who it is for

EurekaClaw is a multi-agent research assistant designed for theoretical research. Its workflow starts with a question or hypothesis typed in the terminal or a browser interface, and ends with a LaTeX paper draft saved locally. Between those points, EurekaClaw crawls arXiv and Semantic Scholar to fetch and summarize relevant papers, generates hypotheses by synthesizing patterns across those papers, produces and verifies formal proofs through a 7-stage bottom-up pipeline, runs numerical experiments to validate theoretical bounds, and flags low-confidence lemmas.

The pyproject.toml describes the project as a multi-agent system for proof-heavy, formalism-rich, math-dense domains. It requires Python 3.11 or later and is licensed under Apache 2.0. Version 0.2.0, released in April 2026, added Paper Q&A, a Rebuttal Helper for responding to reviewer comments, and a Paper Rewrite feature. The last push to the repository was on 2026-06-13.

The 7-stage theorem-proving pipeline

The README describes the theorem prover as a 7-stage bottom-up pipeline that generates, verifies, and formalizes proofs. The stages are not individually named in the README, but the feature table describes the outcome: theorems are generated, stress-tested as hypotheses, and formalized into LaTeX theorem environments with accompanying citations.

The experiment runner complements the proof pipeline by numerically validating theoretical bounds. When a theoretical claim (such as an asymptotic complexity bound) can be checked numerically, EurekaClaw runs that check and flags any lemma whose confidence is low. The README's terminal session demo shows the pipeline producing a theorem from a sparse attention efficiency query, generating a hypothesis about O(n log n) via topological filtration, drafting the theorem, completing the proof, and saving the paper to ./results/.

Continual learning is listed as a separate feature: EurekaClaw distills proof strategies into skills after every session, so performance improves over time as session history accumulates.

Installing EurekaClaw on macOS, Linux, and Windows

On macOS and Linux, a single installer command clones the repository, creates a virtual environment, installs EurekaClaw, and adds the eurekaclaw command to PATH:

curl -fsSL https://eurekaclaw.ai/install.sh | bash

On Windows:

powershell -c "irm https://eurekaclaw.ai/install_win.ps1 | iex"

After installation, the README instructs running `eurekaclaw onboard` to configure the API key and settings. The LLM backend and API key are set in the environment file. The .env.example in the repository is the starting point; the key variables are LLM_BACKEND (defaulting to anthropic) and ANTHROPIC_API_KEY. Copying the example file before editing it preserves the original:

cp .env.example .env

The Docker path is the recommended option for servers. The pre-built CPU image is approximately 10 GB and requires only Docker with the user in the docker group, no Python or Node.js on the host.

CLI commands and Docker browser UI

The CLI exposes the research workflow as direct commands. To prove a theorem claim:

eurekaclaw prove "The sample complexity of transformers is O(L·d·log(d)/ε²)"

To explore a literature area and generate hypotheses:

eurekaclaw explore "multi-armed bandit theory"

For the browser UI, the Docker Compose setup from the repository is the documented path:

docker compose up

This launches the browser UI at http://localhost:8080. The docker-compose.yml defines a production service (eurekaclaw), a GPU service (eurekaclaw-gpu, activated with --profile gpu), and a development service with hot-reload (eurekaclaw-dev, activated with --profile dev). The GPU image is approximately 13 GB and requires an NVIDIA GPU with CUDA 12.4.

The browser UI provides a split-view interface with PDF and LaTeX preview alongside a Q&A chat panel, a live pipeline tracker, and a proof sketch view.

LLM backend options and the .env configuration

EurekaClaw is not tied to a single model provider. The .env.example lists six LLM_BACKEND values. The default is anthropic, which uses the Anthropic API directly with an ANTHROPIC_API_KEY. The openai_compat backend supports any OpenAI-compatible endpoint, covering OpenRouter, vLLM, SGLang, and LM Studio. The openrouter shortcut sets up OpenRouter with an OPENAI_COMPAT_API_KEY. The local shortcut connects to a local vLLM instance at http://localhost:8000/v1. The minimax and novita shortcuts support their respective providers.

For Claude Pro and Max accounts, EurekaClaw can authenticate via OAuth without an API key, using ccproxy. The .env.example documents two OAuth options: a full ccproxy setup using `ccproxy auth login claude_api`, or automatic credential detection from ~/.claude/.credentials.json for users who are already logged in to Claude Code.

The pyproject.toml lists these as optional dependency groups: embeddings (for semantic skill retrieval), docker (for sandboxed code execution), openai, oauth, codex, and pdf (for full PDF extraction via Docling).

Paper Q&A, rebuttal help, and paper rewrite

Version 0.2.0 added two features to the paper workflow. The Paper Q&A and Rebuttal Helper lets users ask multi-turn questions about any generated paper PDF. The README describes this as useful for drafting precise, citation-backed responses to reviewer comments. Reviewers typically ask for clarification on specific lemmas or proofs; the Q&A tool locates the evidence in the paper and formulates a response.

The Paper Rewrite feature revises the paper in one step, taking either a free-form prompt or accumulated Q&A feedback as input. The README notes that paper rewrites include versioned saves and automatic rollback, so earlier drafts are preserved if the revision is not an improvement. These features are documented in the project's user guide at eurekaclaw.github.io, referenced in the README.

Limitations and wrong-fit cases

EurekaClaw's design assumes the research task involves formal proofs and LaTeX papers. Empirical machine learning research, data analysis pipelines, or software engineering work that does not produce theorems and proofs falls outside the documented scope. The pyproject.toml describes the target as proof-heavy, formalism-rich, math-dense domains.

All AI operations require a configured LLM backend. EurekaClaw does not include a built-in model; it routes every generation task through whichever backend is configured. Running the default Anthropic backend requires an API key with token usage costs. Running a local model via the local backend requires a separate vLLM or compatible server.

The repository is at version 0.2.0 with the last push on 2026-06-13. There are no GitHub releases; the distribution uses pre-built Docker images and the install scripts hosted at eurekaclaw.ai. The project format and CLI may change before a stable 1.0 release.

Comparison with using Claude directly for research

The direct alternative to EurekaClaw for researchers using Claude is working through claude.ai or the Anthropic API directly, manually writing prompts for each step: literature search, hypothesis generation, proof drafting, and paper writing. This requires no installation and works with any Claude model.

EurekaClaw adds structure to that workflow: arXiv crawling is automated, the 7-stage proof pipeline validates each step before proceeding to the next, numerical experiment runs check bounds without the researcher setting that up manually, and the continual learning layer retains proof strategies across sessions. The trade-off is the setup overhead: a working Docker environment or a Python 3.11 install, an API key or OAuth configuration, and familiarity with the eurekaclaw prove and eurekaclaw explore commands. For a researcher who writes one paper per year, the direct API approach may be faster to set up. For a researcher running many proofs regularly, the automation and retained skills may reduce the total time per paper.

Editorial conclusion

EurekaClaw suits researchers in mathematics, theoretical machine learning, and adjacent fields who want automated literature crawling, proof generation, and LaTeX drafting without assembling those steps manually. It is a poor fit for empirical research workflows that do not involve formal proofs, or for teams that cannot supply an LLM API key or run a local model server. The last push to the repository was on 2026-06-13. Before starting, set LLM_BACKEND and the appropriate API key in backend/.env, or copy .env.example and run docker compose up for the browser UI.

Frequently asked questions

Official sources

1. EurekaClaw/EurekaClaw on GitHub
2. Issues
3. License: Apache-2.0
4. Project website
5. README