EurekaClaw: An Autonomous Research Pipeline You Install With One curl Command
The official repo of EurekaClaw
At a glance
- What is it?
- EurekaClaw is an Apache-2.0 multi-agent research assistant that crawls arXiv, drafts theorems and proofs, runs numerical experiments, and writes LaTeX papers from a terminal or browser UI. The interesting part is not the feature list but what the pipeline can and cannot verify about its own output.
- Who is it for?
- Adopt EurekaClaw if you want a local-first, Apache-2.0 harness that turns a research question into a draft paper with proof sketches and numerical checks, and you are willing to read every generated lemma yourself. Skip it if you need a citable, peer-reviewed pipeline or a stable release channel, because no releases are listed and the current version is a v0.2.0 README update.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 95 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap EurekaClaw targets: literature search to draft paper in one session
Most LLM research tools stop at retrieval or summarization. EurekaClaw's README positions it at the other end: a single command such as eurekaclaw prove "Find recent papers on sparse attention + prove efficiency bound" is supposed to crawl arXiv, summarize results, generate a hypothesis, draft a theorem with a LaTeX proof, and save a paper draft to ./results/. The stated audience is researchers and engineers who already have a question and want a first draft, not a survey. The README frames the tool as local-first and privacy-oriented, with the browser UI and CLI as equal front ends. That framing matters because it tells you the intended unit of work is a session, not a single API call.
Seven-stage proof pipeline and the agents around it
The README describes the theorem prover as a 7-stage bottom-up pipeline, but it does not enumerate the stages, so the exact decomposition is not verifiable from the supplied material. What is visible is the division of labour: a literature crawler pulls from arXiv and Semantic Scholar, an idea generator synthesizes hypotheses across papers, the prover generates and formalizes proofs, an experiment runner numerically validates theoretical bounds and flags low-confidence lemmas, and a paper writer emits camera-ready LaTeX with theorem environments and citations. A continual learning component is said to distill proof strategies into skills after each session. The experiment runner is the most concrete claim: it is described as flagging low-confidence lemmas, which implies the system has an internal confidence signal rather than treating every generated lemma as equally sound. How that signal is computed is not documented in the README.
Install paths: install.sh, PowerShell, and a 10 GB Docker image
On macOS and Linux the documented install is curl -fsSL https://eurekaclaw.ai/install.sh | bash. The README states this clones the repo, creates a virtual environment, installs EurekaClaw, and adds the eurekaclaw command to PATH. Windows uses powershell -c "irm https://eurekaclaw.ai/install_win.ps1 | iex". After either path you run eurekaclaw onboard to configure the API key and settings. The Docker route is more explicit: docker pull eurekaclaw/eurekaclaw fetches a roughly 10 GB CPU image (Python 3.11, Node.js 18), and the gpu tag is roughly 13 GB with CUDA 12.4. A minimal run is docker run --rm -it -p 8080:8080 -e ANTHROPIC_API_KEY=sk-ant-... eurekaclaw/eurekaclaw, after which the browser UI is at http://localhost:8080. For persistence the README shows --env-file .env plus -v ~/.eurekaclaw:/root/.eurekaclaw. For remote servers it recommends --network host so OAuth callbacks resolve, with SSH port forwarding from the client. Note the Docker path requires the user to be in the docker group; sudo is not needed on the host.
Paper Q&A, rebuttal drafting, and rewrite with rollback
The v0.2.0 update adds two features that change the workflow more than the headline features do. Paper Q&A lets you ask multi-turn questions about any generated paper PDF, which the README frames as a way to draft citation-backed responses to reviewer comments. Paper Rewrite revises the paper from a free-form prompt or accumulated Q&A feedback, and the README states it includes versioned saves and automatic rollback. Rollback is the detail worth noticing: it implies the rewrite step is expected to sometimes make the paper worse, which is a reasonable assumption for any LLM editing loop. Session history is also listed, so completed sessions can be revisited and revised. The README does not describe how versions are stored or how far back rollback can go.
Where the pipeline is likely to mislead you
The honest limitation is stated by the project itself, indirectly: the experiment runner flags low-confidence lemmas, which means some generated lemmas are low confidence. A tool that drafts theorems and proofs will produce plausible-looking LaTeX whether or not the mathematics holds. Nothing in the README describes an external proof checker such as Lean, Coq, or Isabelle, so verification appears to rest on the model plus the numerical experiment runner. Numerical validation of a bound is not a proof of that bound. If your work requires machine-checked proofs, EurekaClaw is the wrong tool. It is also a poor fit if you need a stable, versioned release: the repository lists no releases, and the most recent documented change is a v0.2.0 README update dated April 20, 2026, with the last push to main on 2026-06-13. Treat the main branch as the artifact.
How it differs from a general coding agent with a search tool
A general coding agent with web search and a Python sandbox can also fetch papers and write LaTeX. The difference here is that the stages are named and ordered for research: crawler, idea generator, prover, experiment runner, writer, then Q&A and rewrite. That ordering is the product. A coding agent has no built-in notion of a low-confidence lemma or a proof stage, and no session history scoped to papers. The trade-off is that the research-specific stages are also the least inspectable from the README: the 7-stage pipeline is named but not enumerated, and the continual learning component is described only as distilling proof strategies into skills. If you want a transparent pipeline you can modify stage by stage, EurekaClaw gives you the seams; if you want a general agent, the seams are overhead. Either way, the Apache-2.0 licence means you can read and fork the stages once you have the repo locally, which the installer already clones for you.
Maintenance cost, licence, and what to check before you commit
The Apache-2.0 licence permits commercial use, modification, and redistribution provided you keep the licence and notices intact; this is not legal advice, so read the LICENSE file in the repo. The practical maintenance costs are the 10 GB CPU image (13 GB for GPU), a Python 3.11+ requirement for non-Docker installs, and a model API key that you supply via environment variable or the onboarding step. Continual learning means the tool writes skill state after each session, so the ~/.eurekaclaw volume is not disposable if you want that state to persist. There are no listed releases, so upgrades mean pulling main or pulling a new Docker tag, and the README does not describe a migration path for skill state between versions. Before adopting, verify three things: the contents of install.sh, since it is piped directly to bash; that your chosen model API is accepted by eurekaclaw onboard; and whether the experiment runner's numerical checks are sufficient for the claims you intend to publish. The last one is a judgement only you can make.
Editorial conclusion
Adopt EurekaClaw if you want a local-first, Apache-2.0 harness that turns a research question into a draft paper with proof sketches and numerical checks, and you are willing to read every generated lemma yourself. Skip it if you need a citable, peer-reviewed pipeline or a stable release channel, because no releases are listed and the current version is a v0.2.0 README update. Before installing, verify the Python 3.11+ requirement, confirm the install.sh script contents, and check that the model API key you intend to use is accepted by the onboarding step.
Community notes