Dr. Claw (OpenLAIR/dr-claw): an open source AI research workspace, not the Inspector Gadget villain

What Dr. Claw actually is, and who it is built for

Dr. Claw is a research workspace, not a model. The package description calls it a "Desktop and mobile UI for Claude Code, Cursor CLI, Codex, and OpenRouter with Research Lab", which is the clearest one-line summary available. The README frames the audience as researchers and builders running end-to-end projects, from shaping an idea to experiments to publication-ready output.

The problem it targets is fragmentation. A typical research week involves reading arXiv, tracking HuggingFace releases, writing experiment code, drafting a paper and then answering reviewers. Dr. Claw puts those in one interface with an agent layer on top. The README describes a News Dashboard that aggregates arXiv, HuggingFace Daily Papers plus Models, Datasets and Spaces, GitHub trending, WeChat 公众号 via RSSHub, X and Xiaohongshu, and a Skills Library with more than 100 built-in research skills that agents discover and apply during a task.

The project is model-agnostic by design. The README names Claude Code, Gemini CLI, Codex and OpenRouter, which means you are not buying a model subscription, you are bringing a CLI you already have. That is the strongest part of the pitch and also the source of most of the setup friction, since the workspace inherits whatever those CLIs require.

How the workspace is put together

The repository layout shows a conventional split. There is a server/ directory with routes, middleware, database, utils and constants, a src/ frontend built with Vite, a shared/ directory for code used on both sides, and an electron/ directory for the desktop build. The package main entry is server/index.js, and the bin map exposes two command names, vibelab and dr-claw, both pointing at server/cli.js.

The data flow is local-first. A SQLite database sits behind DATABASE_PATH, and authentication uses JWT tokens signed with JWT_SECRET. If JWT_SECRET is unset, the README states a random secret is generated on first start and stored in a jwt-secret file next to the database. The .env.example documents PORT=3001 for the server and VITE_PORT=5173 for the dev client, with HOST defaulting to 0.0.0.0 and the comment recommending 127.0.0.1 to restrict to localhost only. That default is worth noticing: out of the box the server binds to all interfaces.

Agent execution is delegated. The .env.example lists optional paths for CLAUDE_CLI_PATH, CURSOR_CLI_PATH, GEMINI_CLI_PATH and CODEX_CLI_PATH, and notes that Codex chat turns and model discovery default to the @openai/codex bundled in node_modules while the embedded shell uses the codex on your PATH. Setting CODEX_CLI_PATH routes all of them to one binary. So Dr. Claw is largely an orchestration and interface layer over CLIs you install separately.

Installing Dr. Claw and running a first session

The README gives two entry points. The desktop route is a .dmg or .exe from GitHub Releases, and the zero-setup route is npx. The package is published on npm as dr-claw, so the npx command pulls the published version without a clone.

Start with the quick path:

npx dr-claw

What you should see is the workspace start using the bundled server; the README describes this as zero-setup instant start. If you prefer the terminal, the release notes mention a chat mode:

dr-claw chat

The README says this runs a fully agentic session with any OpenRouter model and needs no browser. Note that the bin map also exposes vibelab as an alias for the same CLI entry point.

For a source install, the scripts in package.json are the reference. The dev script runs a dependency check and then starts the server and Vite client together:

npm install
npm run dev

The server script watches server/routes, server/middleware, server/database, server/utils, server/constants and server/index.js. For a production-style run, npm start builds the frontend and then starts the server. Before any of this, copy .env.example to your environment file and set at least the values you care about:

PORT=3001
VITE_PORT=5173
HOST=127.0.0.1

The .env.example also lists CONTEXT_WINDOW and VITE_CONTEXT_WINDOW, which it says are auto-detected from the model unless you override them. The README does not document a rollback procedure for the desktop installers, so keep the release version you came from if you plan to downgrade.

Where Dr. Claw gets in your way

The model-agnostic design cuts both ways. Because Dr. Claw drives external CLIs, a broken or missing claude, gemini, codex or agent binary is a Dr. Claw problem from the user's perspective even though the fault is upstream. The .env.example exists precisely because those binaries can live in different places, and the Codex note about the bundled node_modules copy versus the PATH copy is the kind of detail that produces confusing behaviour when the two versions differ.

Resource use is another constraint. The README states the workspace runs on your own machine with your own data and GPUs. That is a privacy advantage and a hardware requirement at the same time. There is no hosted tier described, so a laptop without a usable GPU still depends on whatever remote model you point the agent at.

The skills library is large but not curated in the README. It says agents auto-discover and apply skills across ideation, code survey, experiment development and analysis, paper writing, review response and delivery. Auto-discovery means the selection logic lives in the agent, not in a menu you approve. If you need a deterministic pipeline where every step is declared in advance, this is the wrong shape of tool.

Finally, the repository is not archived and the last push was on 2026-09-08, so the codebase is moving. The most recent tagged release is v1.1.4 from 2026-04-12, which means main has run ahead of the release tags by roughly five months. Pin a tag if you need stability.

How it differs from a plain coding agent CLI

The obvious alternative is to use Claude Code, Gemini CLI or Codex directly and skip the workspace. That comparison is fair, and for a single coding task the CLI alone is simpler: no server, no database, no JWT secret, no port. Dr. Claw adds value only when the work spans more than code.

The specific difference is the research lifecycle wrapper. A coding agent has no concept of a news feed, a skills library or a paper draft. Dr. Claw adds a News Dashboard with scored feeds from arXiv, HuggingFace, GitHub, WeChat 公众号, X and Xiaohongshu, plus the Research Lab and Files sidebar tabs described in the release notes. It also adds an Auto Research Hub, which the release notes describe as picking a tool pack (ARIS, Autoresearch, DeepScientist), configuring it, and choosing a workflow in Chat so the agent runs the pipeline from idea to paper.

So the honest split is this: if your bottleneck is writing code, a CLI is enough. If your bottleneck is keeping up with literature and turning findings into a manuscript, the wrapper earns its overhead. The README positions the project against Anthropic's Claude Science workbench, noting that Dr. Claw is open source, model-agnostic and runs locally, which is the axis where a hosted workbench cannot compete on data control.

Maintenance, licence and upgrade cost

Maintenance looks current. The repository is not archived and the last push was on 2026-09-08, close to the time of writing. The release cadence is tight: v1.1.2 on 2026-04-05, v1.1.3 on 2026-04-08 and v1.1.4 on 2026-04-12, three tags in eight days. The CHANGELOG.md and .release-it.json files in the repository root suggest a scripted release process, and release.sh is wired to an npm release script.

The upgrade cost is mostly environmental. Because Dr. Claw shells out to other CLIs, a Dr. Claw upgrade can change which agent behaviour you get, particularly around the bundled Codex copy described in .env.example. The postinstall script runs scripts/fix-node-pty.js, which means native module patching happens on every install; that is a common source of install failures on unusual Node versions. The .nvmrc file in the repository root is the safest place to read the intended Node version.

On licensing, the README badge states GPL-3.0 plus AGPL-3.0, while the repository metadata reports NOASSERTION. Those two do not agree, so treat the LICENSE and NOTICE files as the authoritative source and read them before you build on the code. AGPL-3.0 in particular reaches network-served deployments, which matters if you plan to host Dr. Claw for other people. This is not legal advice; if the distinction affects your product, ask someone qualified.

Editorial conclusion

Adopt Dr. Claw if you already run a coding agent CLI and want news, skills and paper work in one local workspace; skip it if you need a hosted service or a permissive licence, because the project states GPL-3.0 plus AGPL-3.0. Before committing, read docs/configuration.md and .env.example, and decide whether HOST should stay 0.0.0.0 or move to 127.0.0.1.

Frequently asked questions

Official sources

1. Issues
2. OpenLAIR/dr-claw on GitHub
3. Project website
4. README
5. Releases