Dr. Claw: an open source research workspace that wires agents into the whole paper pipeline
A Super AI Lab with massive AI Doctors as Assistants. Best IDE for Research via AI Power.
At a glance
- What is it?
- Dr. Claw is a JavaScript research workspace from OpenLAIR that runs Claude Code, Gemini CLI, Codex or OpenRouter models against a shared project folder, with a news feed, a skills library and an autonomous research hub bolted on. It is model-agnostic and self-hosted, but the README leaves the sandboxing and data-handling story largely unstated.
- Who is it for?
- Adopt Dr. Claw if your group already has working CLI agents, wants the research pipeline in one self-hosted workspace, and can live with a GPL-3.0 plus AGPL-3.0 licence and a README that does not spell out sandboxing. Do not adopt it if you need a hosted service with a support contract, or if you cannot run Node and your own model credentials.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 2 days ago.
- What is it written in?
- Mainly JavaScript, 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 Dr. Claw targets: research work split across a dozen tools
A research project does not live in one program. A survey lives in a reference manager, ideas live in a notes app, experiments live in a repo with its own environment, the paper lives in LaTeX, and the slides live somewhere else again. Each handoff loses context. Dr. Claw's stated aim is to keep that whole chain in one workspace: survey, ideation, experiments, paper writing, then slides and promotion. The README describes it as a general-purpose research assistant for researchers and builders running end-to-end projects.
The target user is narrower than the phrase suggests. You need a working CLI agent already, because Dr. Claw drives Claude Code, Gemini CLI, Codex, or models routed through OpenRouter. The README also positions it against Anthropic's Claude Science workbench, noting that Dr. Claw has been shipping since February 2026, is open source, is model-agnostic, and runs on your own machine with your own data and GPUs. That framing tells you who it is for: people who want the workbench pattern but not a vendor-hosted one. It is not a tool for someone who wants to type a question into a browser and get a paragraph back.
How the pieces fit: agents, a skills library, and a project folder
The mechanism visible in the README is a workspace wrapped around external agent CLIs. You pick a model backend, and Dr. Claw gives the agent a project context plus a set of skills to draw on. The Skills Library is described as 100+ built-in research skills spanning ideation, code survey, experiment development and analysis, paper writing, review response, and delivery. The README says agents auto-discover and apply them as task-level assistance, which implies the skills are files the agent reads rather than functions the app calls. That distinction matters: if skills are prompt-level assets, their behaviour depends heavily on which model backend you selected, and swapping Claude Code for an OpenRouter model may change results without any code change.
Around that core sit three supporting surfaces. The News Dashboard aggregates arXiv, HuggingFace Daily Papers plus Models, Datasets and Spaces, GitHub trending and new repos, WeChat 公众号 via RSSHub, X, and Xiaohongshu into one scored feed you can filter by source and by research domain. The Auto Research Hub lets you pick a tool pack (ARIS, Autoresearch, or DeepScientist), configure it, then choose a workflow in Chat so the agent runs a pipeline from idea to paper. Multi-Tab Sidebar puts Research Lab and Files side by side, and a File Preview Overlay renders project files inline. The data flow is therefore: external feeds and your project files feed the workspace, the agent reads both, and outputs land back in the project folder.
Getting it running: npx, the desktop build, and terminal chat
The README gives three entry points. The fastest is npx dr-claw, described as zero-setup instant start, which requires a Node toolchain but no clone. The second is a native desktop app: .dmg and .exe builds are published on GitHub Releases. The third is terminal-only, dr-claw chat, which the README says runs a fully agentic session with any OpenRouter model and needs no browser.
Configuration is where the real work sits. You choose a model backend, and the README names Claude Code, Gemini CLI, Codex, and OpenRouter as options. For Codex there is a reasoning effort selector, and for Gemini a thinking strength selector, both exposed in the Chat interface. The Auto Research Hub adds a second configuration layer: pick a tool pack from ARIS, Autoresearch, or DeepScientist, configure it, then select a workflow in Chat. Local GPU detection appears in the changelog, which suggests the workspace tries to identify available hardware on startup, though the README excerpt does not say what it does with that information.
One practical note: the README does not list required environment variables or config file paths in the excerpt provided, so the exact keys for model credentials are not verifiable from this material. Check the repository's Configuration section before assuming a single settings file covers all four backends.
Where the design strains: sandboxing, licence, and agent autonomy
The Auto Research Hub is the most consequential feature and the least specified. The README says you can launch fully autonomous research and watch the agent run the entire pipeline from idea to paper while you sleep. An agent with filesystem access, shell access through a CLI backend, and a mandate to run experiments is a broad permission set. The README excerpt does not describe a sandbox, a permission prompt, a dry-run mode, or a rollback mechanism. That is not proof none exists, but it is a gap a prospective user should close before pointing the tool at a machine holding unpublished data or shared credentials.
The licence is the second constraint. The repository's metadata reports NOASSERTION, while the README badge states GPL-3.0 plus AGPL-3.0. Those two signals disagree, and the combination is unusual enough that you should read the LICENSE file directly rather than trust either. If the AGPL component applies to the workspace as a whole, running a modified Dr. Claw as a network service for other people triggers obligations that a purely internal deployment does not. That is a factual difference in obligation, not a judgement about the licence.
Third, the changelog cadence is fast. Three releases landed within a week in April 2026 (v1.1.2, v1.1.3, v1.1.4), and features such as the desktop app, multi-tab sidebar, file preview, terminal chat, reasoning controls and GPU detection all arrived within roughly a month. Fast iteration on a workspace that holds your research projects means upgrade churn is a real cost, and pinning a version before a paper deadline is the sensible default.
What it is not: the case for a plain agent CLI
The honest alternative is the thing Dr. Claw wraps: Claude Code, Gemini CLI, or Codex used directly in a repository. That approach gives you the same model, the same shell access, and the same file editing, with no additional layer and no additional licence. You lose the skills library, the aggregated news feed, the Research Lab sidebar, and the tool-pack launcher. You keep full control over what the agent reads and writes, because you configure it yourself.
The difference in approach is where the research context lives. A bare CLI agent starts each session cold unless you have written your own project instructions and prompt files. Dr. Claw's bet is that a curated set of 100+ skills plus a persistent project workspace beats ad hoc prompting. That bet is testable but not verified here. If your workflow is already a single repository with a well-maintained agent instruction file, the marginal gain from Dr. Claw is mostly the news dashboard and the tool-pack launcher, not the agent itself. If your workflow spans reference management, note-taking, LaTeX, and slides across separate tools, the consolidation argument is stronger.
Maintenance, upgrades, and what the licence implies for deployment
Dr. Claw is JavaScript, distributed on npm as dr-claw, with a desktop build on GitHub Releases. That gives you two upgrade paths with different costs. npx dr-claw pulls the current published version each time, which is convenient and also means you get whatever shipped most recently. The desktop .dmg or .exe is a pinned artifact you replace manually, which is slower but reproducible. For a group running the same pipeline across several people, the desktop build is the easier thing to keep in sync.
The maintenance burden beyond upgrades is the model backends. Dr. Claw depends on Claude Code, Gemini CLI, Codex, and OpenRouter remaining usable in the way it expects. A breaking change in any of those CLIs, or a model retirement on OpenRouter, lands on you rather than on a vendor with a support contract. There is a companion repository, dr-claw-plugin-cc, which the README links as a Claude Code plugin, so at least part of the integration is maintained separately.
On licensing, the README badge says GPL-3.0 plus AGPL-3.0 while the repository metadata says NOASSERTION. Treat the badge as a claim to verify, not a settled fact. If your organisation forbids AGPL code in network-facing services, resolve that discrepancy before you build anything on top of the workspace. This is a description of what the files say, not legal advice.
Who should adopt it, and what to check first
Dr. Claw fits a lab or small team that already runs CLI agents, wants the research pipeline in one self-hosted place, and is comfortable reading source when the documentation runs out. It fits people who want the workbench pattern without sending their data to a hosted service, and who value model choice enough to switch backends mid-project.
It does not fit anyone who needs a hosted product with uptime guarantees, or who cannot run Node and supply their own model credentials. It also does not fit groups whose compliance rules conflict with the AGPL network clause, at least not until the licence question is resolved.
Before adopting, do three concrete things. Run npx dr-claw against a throwaway project folder and confirm which skills load and which model backend you can actually authenticate. Read the LICENSE file in the repository to settle the GPL-3.0 plus AGPL-3.0 versus NOASSERTION discrepancy. And check whether the Auto Research Hub offers any permission or sandbox control before you let it run unattended on a machine that holds unpublished results. The README excerpt reviewed here does not answer that last question, and it is the one that decides whether the autonomy feature is usable in a real lab.
Editorial conclusion
Adopt Dr. Claw if your group already has working CLI agents, wants the research pipeline in one self-hosted workspace, and can live with a GPL-3.0 plus AGPL-3.0 licence and a README that does not spell out sandboxing. Do not adopt it if you need a hosted service with a support contract, or if you cannot run Node and your own model credentials. Before committing, run npx dr-claw on a throwaway project folder, confirm which of the 100+ skills actually load, and read the LICENSE file to check the AGPL network clause against how you plan to deploy it.
Community notes