Dr. Claw: Your AI Research Assistant
Full-stack research workspace.
🎉 Accepted to EMNLP 2026 System Demonstrations
English | 中文
Excited about Anthropic’s Claude Science (launched June 30, 2026)? Dr. Claw has been shipping the same vision since February 2026 — and it’s fully open source, model-agnostic (Claude Code, Gemini CLI, Codex, plus hundreds of models via OpenRouter), and runs on your own machine with your own data and GPUs. It also covers the full research lifecycle — survey → ideation → experiments → paper writing → slides & promotion — not just computational analysis. Free to use; no subscription required.
Table of Contents
Overview
Dr. Claw is a general-purpose AI research assistant designed to help researchers and builders execute end-to-end projects across different domains. From shaping an initial idea to running experiments and preparing publication-ready outputs, Dr. Claw keeps the full workflow in one place so teams can focus on research quality and iteration speed.
Product Screenshot
📰 News Dashboard — A unified, scored feed across arXiv, HuggingFace, GitHub, WeChat 公众号, X, and Xiaohongshu
Filter by source, configure your research domains, and never leave your workspace to keep up with the field — covers arXiv, HuggingFace (Daily Papers + Models / Datasets / Spaces), GitHub trending & new repos, WeChat 公众号 (via RSSHub), X, and Xiaohongshu.
📚 Skills Library — 100+ built-in research skills, auto-discovered and applied by agents
Browse skills spanning ideation, code survey, experiment development & analysis, paper writing, review response, and delivery — agents auto-discover and apply them as task-level assistance throughout the pipeline.
The Philosophy: Leveraged Cognition
Manual work is too slow. Fully automated AI is too generic. Vibe Researching is the new frontier. Dr. Claw turns your Research Taste into outsized outcomes with Agentic Execution–so you can move faster, think bigger, and still hold the line on scientific rigor.
What’s New
- 🧩 v1.3.0: Choose Your Skills, Compact Sidebars
2026-11-06 — Pick which skills are on by default (globally or per project) in a new card-based Skills Library, or choose a preset on first launch. The built-in library is trimmed to 121 working skills, and skills you import now live outside the app so updates keep them. The left sidebar shows one line per project and session and can search inside session messages; the Files panel is a compact tree. The first launch after installing or upgrading also loads reliably now.
- 🧭 v1.2.0: Redesigned Research Workbench
2026-11-05 — A project-centered workspace with a cleaner chat composer and unified controls. The desktop app signs you in automatically, and local models can now run on vLLM and SGLang as well as Ollama. Download it from GitHub Releases.
- 🎉 Accepted to EMNLP 2026
2026-08-22 — Our paper Dr. Claw: An AI Scientist Workspace for Vibe Research has been accepted to the EMNLP 2026 System Demonstrations track! Catch the live demo in Budapest, Hungary, October 24–29, 2026.
- 🧪 Auto Research Hub
2026-04-08 — One click to launch fully autonomous research! Pick a tool pack (ARIS, Autoresearch, DeepScientist), hit configure, choose a workflow in Chat — and watch the agent run your entire research pipeline from idea to paper while you sleep.
- 🖥️ Desktop App & npx
2026-04-06 — Dr. Claw now runs as a native desktop app! Grab the .dmg / .exe from GitHub Releases, or run npx dr-claw for zero-setup instant start.
- 🗂️ Multi-Tab Sidebar
2026-04-06 — Research Lab and Files now live side-by-side as switchable tabs in the right sidebar — everything you need, one glance away.
- 📂 File Preview Overlay
2026-04-06 — Preview any project file inline with a sleek pill toggle and sidebar browser — no more context-switching!
- 💬 Terminal Chat
2026-04-06 — Love the terminal? Run dr-claw chat for a fully agentic session with any OpenRouter model — zero browser required.
- 🎛️ Reasoning Controls
2026-03-30 — Fine-tune your AI’s thinking! Codex reasoning effort and Gemini thinking strength selectors are now right in Chat.
- 🖥️ Local GPU Detection
2026-03-30 — Dr. Claw automatically detects your local GPU resources — ready to put that hardware to work.
- 🌐 OpenRouter Provider
2026-03-28 — Unlock hundreds of models (GPT-5, Claude, Gemini, DeepSeek, Llama, Kimi, and more) with a single API key. The world’s models at your fingertips!
- 🔄 Session Recovery
2026-03-26 — Crashed mid-session? No sweat — hit the retry button and pick up right where you left off.
Earlier updates
- 📡 Auto Port Fallback
2026-03-26 — Port already taken? Dr. Claw finds a free one automatically. One less thing to worry about.
- 🏷️ Smart Prompt Loading
2026-03-26 — Tasks auto-load into Chat with a handy badge dropdown — just click and go!
- 🏷️ Session Stage Tags
2026-03-26 — Sessions are now auto-tagged by research stage — instantly see where each conversation stands.
- ✍️ Rebuttal Skill
2026-03-24 — New skill for crafting review rebuttals — turn reviewer feedback into publication-ready responses.
- 🧑💻 Multi-Session Support
2026-03-21 — Run multiple sessions in parallel with smart naming — juggle projects like a pro!
- 🗑️ Trash Bin
2026-03-21 — Accidentally deleted a project? Relax — it’s in the trash, ready to be restored.
- 🤖 Dr. Claw CLI & OpenClaw
2026-03-21 — Full CLI control plus an OpenClaw integration for mobile-friendly, voice-ready research management.
- 📚 Reference Library
2026-03-20 — Manage your papers with a streamlined picker and local Zotero support — your literature, organized.
- 🔀 Git Source Control
2026-03-20 — Stage, commit, diff, and switch branches without ever leaving the app. Version control, built in.
- 📰 News Dashboard
2026-03-14 — Stay on top of research-relevant updates right inside your workspace — never miss a trending paper!
Highlights
- 🔬 Research Lab — Structured dashboard for end-to-end research: define your brief, generate a pipeline of tasks, track progress across Survey → Ideation → Experiment → Publication → Promotion, and inspect source papers, ideas (rendered with LaTeX math), and cache artifacts — all at a glance
- ⚡ Auto Research — Start one-click sequential task execution directly from the Project Dashboard, open the generated session live, and receive an email when the run completes
- 📚 100+ Research Skills — A curated library spanning idea generation, code survey, experiment development & analysis, paper writing, review response, and delivery — automatically discovered by agents and applied as task-level assistance
- 🗂️ Chat-Driven Pipeline — Describe your research idea in Chat; the agent uses the
inno-pipeline-planner skill to interactively generate a structured research brief and task list — no manual templates needed
- 🤖 Multi-Agent Backend — Seamlessly switch between Claude Code, Gemini CLI, Codex, and OpenRouter as your execution engines
What the Pipeline Produces
|
Artifact |
Location |
Description |
| 📚 |
Survey reports |
Survey/reports/ |
Literature reviews with citations from arXiv, Semantic Scholar, and web sources |
| 💡 |
Research ideas |
Ideation/ideas/ |
Brainstorming outputs with multi-persona evaluation scores |
| 🔬 |
Experiment code |
Experiment/core_code/ |
Implementation from the plan → implement → judge loop |
| 📊 |
Analysis results |
Experiment/analysis/ |
Statistical analysis, tables, and paper-ready figures |
| 📝 |
Paper draft |
Publication/paper/ |
Academic manuscript (IEEE/ACM format) with citations and LaTeX math |
| 🎞️ |
Presentation |
Promotion/slides/ |
Slide deck, TTS narration audio, and demo video |
See docs/pipeline-outputs.md for the full artifact list and project directory structure.
More Features
- 💬 Interactive Chat + Shell — Chat with your agent or drop into a full terminal — side by side with your research context
- 📁 File & Git Explorer — Browse files with syntax highlighting, live-edit, stage changes, commit, and switch branches without leaving the UI
- 📱 Responsive & PWA-Ready — Desktop, tablet, and mobile layouts with bottom tab bar, swipe gestures, and Add-to-Home-Screen support
- 🔄 Session Management — Resume conversations, manage multiple sessions, and track full history across projects
Feature Gallery
Expand screenshots
Project Dashboard — Start from the project overview, review status, and launch end-to-end automation.
Skill Library — Browse reusable research skills across ideation, experimentation, and writing.
News Dashboard — Follow research-relevant updates without leaving the workspace.
Quick Start
Recommended: Desktop App
The easiest way to use Dr. Claw is the desktop app. It runs the server for you, so there is no terminal or browser setup, and it signs you in automatically.
- Download the installer for your platform from the latest release:
- macOS (Apple Silicon):
Dr.Claw-<version>-mac-arm64.dmg
- Windows (x64):
Dr.Claw-<version>-win-x64.exe
- Install and open it.
- Follow the first-run setup to connect at least one agent (Claude Code, Codex or Gemini). You can change this later under Settings → Agents.
Intel Macs and Linux are not covered by the installers yet. Use a source install (Option 2) for the latest version; npx (Option 1) also works but may install an older release.
Prerequisites
These are for Option 1 and Option 2 below. The desktop app bundles its own runtime, so it doesn’t need Node.js or build tools.
- Node.js v20 or higher (v22 LTS recommended, see
.nvmrc)
- At least one of the following CLI tools installed and configured:
- Some systems need native build tools for dependencies like
node-pty and better-sqlite3. If npm install fails, see FAQ.
Cursor agent support is in progress and coming soon.
Option 1: Run with npx
No cloning or manual setup required — just run:
npx dr-claw
Or install globally for repeated use:
npm install -g dr-claw
dr-claw
Then open your browser at http://localhost:3001 to create your account and start using Dr. Claw.
Note: the npm package can lag behind the desktop release. Compare npm view dr-claw version with the latest release; if npm is older, use the desktop app or a source install to get the newest features.
Option 2: Install from Source (For Development)
- Clone the repository:
git clone https://github.com/OpenLAIR/dr-claw.git
cd dr-claw
- Install dependencies:
npm install
- Configure environment:
cp .env.example .env
# Edit .env with your preferred settings (port, etc.)
Need custom ports, auth, or workspace settings? See docs/configuration.md.
- Start the application:
# Development mode (with hot reload)
npm run dev
Then create your account via the browser http://localhost:5173.
- Use the application
There are two ways to interact with Dr. Claw: the frontend UI workflow or the terminal-only. The UI provides richer visualization but may encounter occasional bugs; the terminal approach is more stable and lightweight.
Option A: Frontend UI
Open the web UI in your browser
Open your browser at http://localhost:5173 (or the port you configured in .env).
Option B: Terminal Only
Use the CLI harness with your preferred agent
Open a second terminal (keep npm run dev running in the first) and install the drclaw CLI harness:
pip install -e ./agent-harness
Then log in with the credentials you created during setup:
drclaw auth login --username YOUR_USERNAME --password YOUR_PASSWORD
Install at least one agent CLI (if you haven’t already):
| Agent |
Install |
Auth |
| Claude Code |
npm install -g @anthropic-ai/[email protected] |
claude → follow OAuth prompt |
| Gemini CLI |
npm install -g @google/[email protected] |
gemini → Google sign-in, or export GOOGLE_API_KEY=... |
| Codex CLI |
npm install -g @openai/codex |
codex login, or export OPENAI_API_KEY=... |
| OpenRouter |
No CLI needed |
export OPENROUTER_API_KEY=sk-or-... (get a key at openrouter.ai/keys) |
OpenRouter lets you use any model (GPT-5, Claude, Gemini, DeepSeek, Llama, Mistral, Qwen, etc.) through a single API key. Select your model in the UI or set OPENROUTER_MODEL in .env.
Navigate to the project directory you want to work in and launch any of the agents:
cd /path/to/your/project
claude # or: gemini | codex
Skills from dr-claw/skills/ are automatically symlinked into each project’s .claude/skills/ directory when the project is created, so the agent discovers them without extra configuration. You can also reference any skill manually inside a session:
> Read .claude/skills/inno-experiment-analysis/SKILL.md and follow it to analyze my results.
Option C: OpenRouter Terminal Chat
Lightweight terminal chat with any OpenRouter model
For a lightweight terminal-only experience using any OpenRouter model, use the built-in dr-claw chat command. No browser or UI required — just an interactive agentic session with full tool-calling capabilities (file I/O, shell, grep, glob, web search/fetch).
# Make sure OPENROUTER_API_KEY is set (or pass --key)
export OPENROUTER_API_KEY=sk-or-...
# Launch a chat session with any model
node server/cli.js chat --model moonshotai/kimi-k2.5
You can also pass the API key inline:
node server/cli.js chat --model anthropic/claude-sonnet-4 --key sk-or-your-key
| Flag |
Description |
--model <slug> |
OpenRouter model slug (e.g., moonshotai/kimi-k2.5, anthropic/claude-sonnet-4, deepseek/deepseek-r1) |
--key <key> |
OpenRouter API key (defaults to OPENROUTER_API_KEY env var). Prefer the env var over --key to avoid exposing secrets in shell history. |
Browse all available models at openrouter.ai/models.
Option D: Desktop App
Run as a standalone desktop application (macOS & Windows)
# Development mode (launches Electron with hot reload)
npm run desktop:dev
# Build distributable installer (.dmg / .exe)
npm run desktop:dist
For details on the desktop architecture, IPC bridge, and CI/CD release process, see electron/README.md.
If agent web search does not work later, see Troubleshooting Web Search below.
For optional web search and page fetching in Claude sessions without a search API
key, see Parallel Search MCP setup.
To publish research summaries or HTML reports from Claude sessions as shareable
links without an API key, see deed.page MCP setup.
OpenClaw Integration
Turn Dr. Claw into a mobile-ready, voice-friendly research secretary
OpenClaw connects to Dr. Claw through the drclaw CLI, giving you project control, smart digests, and proactive notifications — all from your phone or chat app.
Architecture
┌─────────────────────────────────────────────────────────────┐
│ User (mobile / chat / voice) │
│ ↕ │
│ OpenClaw ── secretary layer ──────────────────────────┐ │
│ │ runs local `drclaw ...` receives push msgs │ │
│ ↓ ↑ │ │
│ drclaw CLI ── stable control plane ──────────────┐ │ │
│ │ JSON + openclaw.* schema WebSocket │ │ │
│ ↓ │ │ │ │
│ Dr. Claw Server Watcher ────┘ │ │
│ (projects, sessions, pipelines, artifacts) │ │
└─────────────────────────────────────────────────────────────┘
The integration has three layers:
| Layer |
What it does |
| Control plane |
OpenClaw executes drclaw --json ... commands locally |
| Structured contract |
JSON responses carry a versioned openclaw.* schema payload |
| Proactive delivery |
An event-driven watcher pushes important changes to Feishu / Lark |
Quick Start (6 steps)
Prerequisites
- Dr. Claw server running locally (
npm run dev or drclaw server on)
- At least one project and one execution backend (Claude Code, Gemini CLI, or Codex)
- OpenClaw with local shell /
exec capability
- (Optional) Feishu / Lark channel access for push notifications
1. Start the server
npm install && npm run dev # or: drclaw server on
drclaw --json auth status # verify reachability
drclaw server status only reports the daemon from drclaw server on. If you started Dr. Claw with npm run dev, it may show STOPPED even though http://localhost:3001 is working — use auth status as the real check.
2. Install the CLI
pip install -e ./agent-harness
drclaw --help
If drclaw is not on your PATH:
PYTHONPATH=agent-harness python3 -m cli_anything.drclaw.drclaw_cli --help
3. Authenticate
drclaw auth login --username <user> --password <pass>
drclaw --json projects list # should return your projects
4. Link OpenClaw
drclaw install --server-url http://localhost:3001
# with push channel:
drclaw install --server-url http://localhost:3001 --push-channel feishu:<chat_id>
This copies the Dr. Claw skill, installs wrapper scripts, and saves the server URL and CLI path.
5. Verify the core loop
Run these four commands from OpenClaw — if they all return valid JSON, the integration is live:
drclaw --json projects list # resolve projects
drclaw --json chat waiting # find sessions needing input
drclaw --json digest portfolio # cross-project summary
drclaw --json workflow status --project <project> # single-project status
6. Reply into a session
drclaw --json chat waiting # pick a session
drclaw --json chat reply --project <proj> --session <sid> -m “Continue with option B.”
drclaw --json chat waiting --project <proj> # confirm it cleared
For multi-turn discussion within the same project:
drclaw --json chat project --project <proj> --session <sid> -m “Summarize blockers.”
Structured Schema
Machine-facing commands return a versioned openclaw field. Current families:
| Schema |
Purpose |
openclaw.turn.v1 |
Single chat turn summary |
openclaw.project.v1 |
Project digest with status, counts, and next actions |
openclaw.portfolio.v1 |
Cross-project overview with recommendations |
openclaw.daily.v1 |
Daily digest |
openclaw.report.v1 |
Mobile-ready report payload |
openclaw.event.v1 |
Watcher event with derived signals |
Client rendering tips:
| When you need to… |
Read this field |
| Decide whether to interrupt the user |
openclaw.decision.needed |
| Show quick actions or voice suggestions |
openclaw.next_actions |
| Render a compact summary |
openclaw.turn.summary or openclaw.focus |
| Handle watcher notifications |
openclaw.event.v1.event.signals |
Always prefer the openclaw payload over raw reply text when both are present.
Full contract: agent-harness/cli_anything/drclaw/SCHEMA.md
Proactive Watcher
The watcher is event-driven — it subscribes to Dr. Claw WebSocket events and only notifies on attention-worthy changes.
# Configure push channel
drclaw openclaw configure --push-channel feishu:<chat_id>
# Manage the watcher
drclaw --json openclaw-watch on --to feishu:<chat_id>
drclaw --json openclaw-watch status
drclaw --json openclaw-watch off
How it works:
WebSocket event → project resolution → snapshot diff → signal derivation
↓
dedup (6h TTL) ← stable signature + signal kinds
↓
openclaw agent --deliver → Feishu / Lark summary
(fallback: plain bridge push)
Derived signals:
| Signal |
Meaning |
human_decision_needed |
Agent requests permission for a tool call |
waiting_for_human |
Session is blocked on user input |
blocker_detected |
A task transitioned to blocked state |
blocker_cleared |
A previously blocked task is now unblocked |
task_completed |
One or more tasks finished |
next_task_changed |
The recommended next task has changed |
attention_needed |
General attention signal |
session_aborted |
A session execution was aborted |
State and logs:
~/.drclaw/openclaw-watcher-state.json~/.drclaw/logs/openclaw-watcher.log
Serialized Local Turns
When OpenClaw calls openclaw agent --local repeatedly, use the wrapper script to avoid session-lock collisions:
agent-harness/skills/dr-claw/scripts/openclaw_drclaw_turn.sh \
--json -m “Use your exec tool to run \`drclaw --json digest portfolio\`. Return only raw stdout.”
Rule of thumb: stable serial turns are always better than risky parallel turns.
Integration Checklist
Your OpenClaw integration is complete when all of these work:
At that point, OpenClaw becomes Dr. Claw’s mobile secretary. Users can speak naturally:
”Which projects are waiting for my reply?”
”Summarize this project’s progress and blockers.”
”Reply to that session: go with option B and report back.”
”Give me a cross-project summary and what to focus on today.”
”I have a new idea — create a project, discuss it with me, and start planning.”
Configuration
Dr. Claw reads local settings from .env. For most users, the only required step is copying .env.example to .env, but these are the settings you are most likely to adjust early:
PORT: backend server portVITE_PORT: frontend dev server portHOST: bind address for the frontend and backendJWT_SECRET: optional; a random secret is generated and stored next to the database on first start. Set it only when you need to control the valueWORKSPACES_ROOT: default root for new project workspaces
For the full environment reference and deployment notes, see docs/configuration.md.
Auto Research email notifications are configured inside the app at Settings → Email. The v1 flow supports Claude Code, Codex, Gemini, and OpenRouter engines for unattended task execution, and interrupted runs are automatically reconciled so they do not remain stuck in running.
OpenRouter Setup
OpenRouter is integrated as a first-class provider, giving you access to hundreds of models (GPT-5, Claude, Gemini, DeepSeek, Llama, Mistral, Qwen, Kimi, and more) through a single API key.
Setup
- Get an API key at openrouter.ai/keys.
- Set the key in one of three ways:
- Environment variable:
export OPENROUTER_API_KEY=sk-or-...
.env file: add OPENROUTER_API_KEY=sk-or-... to your project .env- UI: go to Settings → OpenRouter and paste your key
Using OpenRouter in the UI
- Open a project and go to Chat.
- Under Choose Your AI Assistant, click OpenRouter.
- Search for a model in the dropdown (it fetches the full list from OpenRouter) or type a custom model slug.
- Start chatting — the agent has the same tool-calling capabilities as Claude, Gemini, and Codex (file read/write, shell, grep, glob, web search/fetch, todo).
OpenRouter is also available in Auto Research on the Project Dashboard — select it as the provider and pick any model.
Using OpenRouter in the Terminal
No browser needed. The dr-claw chat CLI gives you a fully agentic terminal session:
# Basic usage
node server/cli.js chat --model moonshotai/kimi-k2.5
# With an explicit API key
node server/cli.js chat --model deepseek/deepseek-r1 --key sk-or-your-key
The CLI supports the same tools as the UI (file I/O, shell, grep, glob, web search, web fetch, todo). Type your message and the agent will execute multi-step research tasks autonomously.
Default Model
Set OPENROUTER_MODEL in .env to change the default model used when none is specified:
OPENROUTER_MODEL=moonshotai/kimi-k2.5
If unset, the default is anthropic/claude-sonnet-4.
Research Lab — Quick Example
The core feature of Dr. Claw is the Research Lab.
Research Lab Screenshot
The typical flow is:
- Configure one supported agent in Settings.
- Configure notification settings in Settings → Email if you want completion email notifications.
- Describe your research idea in Chat.
- Let the agent generate
.pipeline/docs/research_brief.json and .pipeline/tasks/tasks.json.
- Review the pipeline in Research Lab and either send tasks back to Chat manually or click Auto Research on the Project Dashboard to run them sequentially.
For full step-by-step operations, see Usage Guide below.
Usage Guide
After starting Dr. Claw, open your browser and follow the steps below.
Step 1 — Create or Open a Project
When you first open Dr. Claw you will see the Projects sidebar. You have two options:
- Open an existing project — Dr. Claw auto-discovers registered projects and linked sessions from Claude Code, Codex, and Gemini.
- Create a new project — Click the “+” button, choose a directory on your machine, and Dr. Claw will set up the workspace: agent folders such as
.claude/, .agents/, .gemini/, standard workspace metadata, linked skills/ directories, preset research dirs (Survey/references, Survey/reports, Ideation/ideas, Ideation/references, Experiment/code_references, Experiment/datasets, Experiment/core_code, Experiment/analysis, Publication/paper, Promotion/homepage, Promotion/slides, Promotion/audio, Promotion/video), and instance.json at the project root with absolute paths for those directories. Cursor agent support is coming soon.
Default project storage path: New projects are stored under ~/dr-claw by default. You can change this in Settings → Appearance → Default Project Path, or set the WORKSPACES_ROOT environment variable. The setting is persisted in ~/.claude/project-config.json.
Step 2 — Generate Your Research Pipeline via Chat
After creating or opening a project, Dr. Claw opens Chat by default. If no research pipeline exists yet, an onboarding banner appears with a Use in Chat button that injects a starter prompt.
Chat Screenshot
Describe your research idea — even a rough one is fine. The agent uses the inno-pipeline-planner skill to ask clarifying questions and then generates:
.pipeline/docs/research_brief.json (your structured research brief).pipeline/tasks/tasks.json (the task pipeline)
Step 3 — Review in Research Lab and Execute Tasks
Switch to Research Lab to review the generated tasks, progress metrics, and artifacts. Then execute tasks:
Task Execution Screenshot
- Choose a CLI backend from the CLI selector (Claude Code, Gemini CLI, or Codex).
- In Research Lab, click Go to Chat or Use in Chat on a pending task.
- The agent executes the task and writes results back to the project.
Optional — Run Auto Research from the Project Dashboard
If you want Dr. Claw to execute the generated task list end-to-end for you, use Auto Research:
- Open Settings → Email and configure
Notification Email, Sender Email, and Resend API Key.
- Make sure your project already contains
.pipeline/docs/research_brief.json and .pipeline/tasks/tasks.json.
- Open the Project Dashboard and click Auto Research on the project card.
- Use Open Session to jump into the live Claude session created for the run.
- When all tasks finish, Dr. Claw sends a completion email. If the session is interrupted, stale runs are recovered automatically so they can be cancelled cleanly instead of staying stuck in
running.
Optional — Configure Auto Research Hub
Auto Research Hub (sidebar → Auto Research) lets you browse and configure third-party research tool packs. Three packs are available:
| Pack |
Description |
Dependencies |
| ARIS |
End-to-end autonomous pipeline with cross-model adversarial review |
MCP Reviewer required |
| Autoresearch |
Goal-directed iteration engine with 9 subcommands |
Zero dependencies |
| DeepScientist |
13-stage research OS (ICLR 2026) with 50+ templates |
Zero dependencies |
How to use:
- Open Auto Research from the sidebar.
- Pick a pack and expand Included Workflows to see available slash commands.
- If the pack requires configuration (e.g. ARIS), expand Configuration, select an MCP Reviewer, enter your API key, and click Auto Configure.
- Go to Chat, click the Auto Research dropdown above the input box, select the pack and workflow.
- Enter your research topic — the agent executes the workflow end-to-end and writes results to your project.
Zero-dependency packs (Autoresearch, DeepScientist) work out of the box — no configuration step needed.
Step 4 — Troubleshooting Web Search
If the agent cannot search webpages, your current permission settings are likely too restrictive. Also check whether a runtime network lock is still active for the process.
- Check the runtime network lock:
echo "${CODEX_SANDBOX_NETWORK_DISABLED:-0}"
If the output is 1, network requests can remain blocked even if Settings permissions are opened. Remove or override this variable in your deployment or startup layer (shell profile, systemd, Docker, PM2), then restart Dr. Claw.
- Open Settings (gear icon in sidebar).
- Go to Permissions, then choose your current agent:
- Claude Code:
- Enable
WebSearch and WebFetch in Allowed Tools.
- Ensure they are not present in Blocked Tools.
- Optionally enable Skip permission prompts if you want fewer confirmations.
- Gemini CLI:
- Choose an appropriate Permission Mode.
- Allow
google_web_search and web_fetch in Allowed Tools when web access is required.
- Ensure they are not present in Blocked Tools.
- Codex:
- In Permission Mode, switch to Bypass Permissions when web access is required.
- Return to Chat, start a new message, and retry your web-search prompt.
Codex permission mode notes:
- Default / Accept Edits: sandboxed execution; network may still be restricted by session policy.
- Bypass Permissions:
sandboxMode=danger-full-access with full disk and network access.
Security note:
- Use permissive settings only in trusted projects/environments.
- After finishing web search tasks, switch back to safer settings.
Step 5 — Resolve "Workspace Trust" or First-Run Errors
Each agent may require a one-time trust confirmation before it can execute code in your project directory. If Chat freezes or shows a trust prompt, switch to the Shell tab inside Dr. Claw and approve the prompt there.
Steps:
- Switch to the Shell tab in Dr. Claw.
- Approve the trust/auth prompt shown in Shell.
- Return to Chat and resend your message.
By default, trust flow is already enabled in Dr. Claw, so you usually do not need to manually run extra trust commands.
The trust decision is persisted per directory — you only need to do this once per project.
Shell tab not working? If the Shell tab shows Error: posix_spawnp failed, see docs/faq.md for the fix, then retry.
You can switch tabs at any time:
| Tab |
What it does |
| Chat |
Start here. Use it to describe your research idea, generate a pipeline, and run tasks with the selected agent. |
| Survey |
Review papers, literature graphs, notes, and survey-stage tasks for the current project. |
| Research Lab |
Review the research brief, task list, progress, and generated artifacts in one place. |
| Skills |
Browse installed skills, inspect their contents, and import additional local skills. |
| Compute |
Manage compute resources and run experiment workloads from one place. |
| Shell |
Use the embedded terminal when you need direct CLI access, trust prompts, or manual commands. |
| Files |
Browse, open, create, rename, and edit project files with syntax highlighting. |
| Git |
Inspect diffs, stage changes, commit, and switch branches without leaving the app. |
Research Skills
Dr. Claw now uses the generated Pipeline Task List as the execution flow.
The project includes 100+ skills under skills/ to support research tasks (idea exploration, code survey, experiment development/analysis, writing, review, and delivery).
These skills are discovered by the agent and can be applied as task-level assistance throughout the workflow.
Claude Code Plugin
If you use Claude Code as your primary coding agent and want Dr. Claw’s research pipeline directly in your terminal — without running the full web UI — check out the standalone Claude Code plugin:
OpenLAIR/dr-claw-plugin-cc
The plugin provides 4 slash commands (/drclaw:setup, /drclaw:status, /drclaw:run, /drclaw:reset) that let you initialize research projects, track progress across all 5 pipeline stages, and execute tasks — all from within a Claude Code session.
Install
Inside any Claude Code session, run:
/plugin marketplace add OpenLAIR/dr-claw-plugin-cc
/plugin install dr-claw@dr-claw
/reload-plugins
The plugin is installed and will be available in future sessions. Run /drclaw:setup to initialize a new research project.
Scope options: By default the plugin is installed to your user scope (available in all projects). Add --scope project to install it for a specific project only, or --scope local for a machine-local install.
What’s Included
- 60+ bundled skills — a curated subset of Dr. Claw’s skill library covering literature survey, idea generation, experiment development, paper writing, and more
- 3 project templates — Method/Model, Dataset/Benchmark, and Position Paper, each with a pre-configured task pipeline
- Auto-detection — the plugin detects existing pipeline projects on session start and shows your current progress
- Same data format — projects created with the plugin use the same
research_brief.json and tasks.json schemas as the full Dr. Claw workspace, so you can switch between them
Note: The plugin bundles a snapshot of skills from this repository. Skills are synchronized manually — see the plugin repo’s README for details.
Additional Details
Mobile, architecture, and security notes
Mobile & Tablet
Dr. Claw is fully responsive. On mobile devices:
- Bottom tab bar for thumb-friendly navigation
- Swipe gestures and touch-optimized controls
- Add to Home Screen to use it as a PWA (Progressive Web App)
Architecture
System Overview
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Frontend │ │ Backend │ │ Agent │
│ (React/Vite) │◄──►│ (Express/WS) │◄──►│ Integration │
│ │ │ │ │ │
└─────────────────┘ └─────────────────┘ └─────────────────┘
Backend (Node.js + Express)
- Express Server - RESTful API with static file serving
- WebSocket Server - Communication for chats and project refresh
- Agent Integration (Claude Code, Gemini CLI, Codex, OpenRouter) - Process spawning, streaming, and session management
- File System API - Exposing file browser for projects
Frontend (React + Vite)
- React 18 - Modern component architecture with hooks
- CodeMirror - Advanced code editor with syntax highlighting
🔒 Important Notice: Agent permissions are configurable per provider. Review Settings → Permissions before enabling broad file, shell, or web access.
To use web and tool-heavy workflows safely:
- Open Settings - Click the gear icon in the sidebar
- Choose an Agent - Claude Code, Gemini CLI, or Codex
- Enable Selectively - Turn on only the tools or permission mode you need
- Apply Settings - Your preferences are saved locally
Recommended approach: Start with the safest permission mode that still lets you complete the task, then relax settings only when needed.
Contributing
Show details
We welcome contributions! Please follow these guidelines:
Getting Started
- Fork the repository
- Clone your fork:
git clone <your-fork-url>
- Install dependencies:
npm install
- Create a feature branch:
git checkout -b feature/amazing-feature
Development Process
- Make your changes following the existing code style
- Test thoroughly - ensure all features work correctly
- Run quality checks:
npm run typecheck && npm run build
- Commit with descriptive messages following Conventional Commits
- Push to your branch:
git push origin feature/amazing-feature
- Submit a Pull Request with:
- Clear description of changes
- Screenshots for UI changes
- Test results if applicable
What to Contribute
- Bug fixes - Help us improve stability
- New features - Enhance functionality (discuss in issues first)
- Documentation - Improve guides and API docs
- UI/UX improvements - Better user experience
- Performance optimizations - Make it faster
For setup help and troubleshooting, see FAQ.
Legacy Compatibility & Deprecation
Dr. Claw was previously known as VibeLab. For users migrating from VibeLab, we provide a compatibility layer during the transition phase:
- CLI Alias: The
vibelab command is still supported as an alias for drclaw but will issue a deprecation warning.
- Python Package: The
VibeLab class in the agent-harness is deprecated; please use the DrClaw class instead.
- Session Files: The CLI now defaults to
~/.drclaw_session.json but will automatically check for and migrate ~/.vibelab_session.json if found.
- Environment Variables:
DRCLAW_URL and DRCLAW_TOKEN are preferred, but VIBELAB_URL and VIBELAB_TOKEN are still supported as fallbacks.
Timeline: We plan to remove legacy vibelab support in Version 2.0 (estimated Q3 2026). Please update your scripts and integrations as soon as possible.
License
This repository contains a combined work.
Upstream portions derived from Claude Code UI remain under GNU General Public License v3.0 (GPL-3.0), while original modifications and additions by Dr. Claw Contributors are licensed under GNU Affero General Public License v3.0 (AGPL-3.0).
See LICENSE and NOTICE for the full license texts and scope details.
Citation
Dr. Claw is described in our paper Dr. Claw: An AI Scientist Workspace for Vibe Research (arXiv:2609.00365), accepted to EMNLP 2026 System Demonstrations. If you find Dr. Claw useful in your research, please cite:
@misc{song2026drclaw,
title = {Dr. Claw: An AI Scientist Workspace for Vibe Research},
author = {Dingjie Song and Hanrong Zhang and Dawei Liu and Yixin Liu and Zongxia Li and Zhengqing Yuan and Siqi Zhang and Henry Peng Zou and Zhiling Yan and Yuxuan Zhang and Yanfang Ye and Philip S. Yu and Lichao Sun},
year = {2026},
eprint = {2609.00365},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
doi = {10.48550/arXiv.2609.00365},
url = {https://arxiv.org/abs/2609.00365},
note = {Accepted to EMNLP 2026 System Demonstrations},
}
Acknowledgments
Built With
Also Thanks To
- Claude Code UI — Dr. Claw is based on it. See NOTICE for details.
- AI Researcher (HKUDS) — Inspiration for research workflow and agentic research.
- Vibe-Scholar — Inspiration for the AI-native research workspace direction.
- autoresearch — Inspiration for autonomous research orchestration and end-to-end execution.
- Autoresearch (uditgoenka) — Autonomous goal-directed iteration engine, integrated as a tool pack in Auto Research Hub.
- ARIS — End-to-end autonomous research pipeline with cross-model adversarial review, integrated as a tool pack in Auto Research Hub.
- DeepScientist — 13-stage autonomous research OS from ICLR 2026, integrated as a tool pack in Auto Research Hub.
Stay Updated
- Star this repository to show support
- Watch for updates and new releases
- Follow the project for announcements
Dr. Claw — From idea to paper.
评论