ResearchClaw

Local-first Research OS for papers, workflows, experiments, channels, and automation.

English | 中文

What It Is

ResearchClaw is not just a chat wrapper. The current codebase is a local-first FastAPI application that combines:

• a long-running app runtime with control-plane APIs
• a web console for chat, papers, research, channels, sessions, cron jobs, models, skills, workspace, environments, and MCP
• multi-agent routing with per-agent workspaces and binding rules
• a persistent research state layer for projects, workflows, tasks, notes, claims, evidences, experiments, artifacts, and drafts
• built-in channels for console, telegram, discord, dingtalk, feishu, imessage, qq, and voice
• model/provider management with multiple providers, multiple models per provider, and fallback chains
• standard SKILL.md support, Skills Hub search/install APIs, MCP client management, and custom channels
• automation triggers, cron jobs, heartbeat, proactive reminders, and runtime observability
• paper search/download, BibTeX utilities, LaTeX helpers, data analysis, browser/file tools, and structured research memory

It is still an Alpha project, but it is no longer just a platform shell. The code now includes a minimal research workflow runtime, claim/evidence graph, experiment tracking, blocker remediation, and project dashboard. The biggest remaining gaps are evidence-matrix quality, stronger claim-evidence validation, richer external execution adapters, and submission/reproducibility packaging.

Quick Start

1) Clone and install from source

git clone https://github.com/MingxinYang/ResearchClaw.git
cd ResearchClaw
pip install -e .

2) Initialize the workspace

researchclaw init --defaults --accept-security

This creates:

• working dir: ~/.researchclaw
• secret dir: ~/.researchclaw.secret
• bootstrap Markdown files such as SOUL.md, AGENTS.md, PROFILE.md, and HEARTBEAT.md

3) Configure a model provider

researchclaw models config

Or add one directly:

researchclaw models add openai --type openai --model gpt-5 --api-key sk-...

Supported provider types in code today: openai, anthropic, gemini, ollama, dashscope, deepseek, minimax, other, custom.

4) Start the service

researchclaw app --host 127.0.0.1 --port 8088

Open http://127.0.0.1:8088.

If the page says Console not found, build the frontend once:

cd console
npm install
npm run build

The backend automatically serves console/dist when it exists.

5) Open the Research page

After startup, open the Research page in the console to:

• create a project
• inspect workflows, claims, and reminders
• view execution health and recent blockers
• dispatch, execute, or resume remediation work

What You Get Today

Runtime and control plane

• FastAPI app with /api/health, /api/version, /api/control/*, /api/automation/*, /api/providers, /api/skills, /api/mcp, /api/workspace, and more
• gateway-style runtime bootstrapping for runner, channels, cron, MCP, automation store, and config watcher
• runtime status snapshots for agents, sessions, channels, cron, heartbeat, skills, automation runs, and research services

Research OS core

• project abstraction with persistent project -> workflow -> task -> artifact relationships
• workflow stages for literature_search, paper_reading, note_synthesis, hypothesis_queue, experiment_plan, experiment_run, result_analysis, writing_tasks, and review_and_followup
• structured notes including paper notes, idea notes, experiment notes, writing notes, and decision logs
• claim/evidence graph that can link papers, notes, experiments, PDF chunks, citations, generated tables, and artifacts
• experiment tracking with execution bindings, heartbeat/result ingestion, contract validation, result bundle validation, and compare APIs
• proactive workflow reminders plus remediation tasks for missing metrics, outputs, or artifact types
• project dashboards and blocker panels, including batch dispatch/execute/resume actions in the console and APIs

Research tools and skills

Built-in tools registered by the agent include:

• semantic_scholar_search
• bibtex_search, bibtex_add_entry, bibtex_export
• latex_template, latex_compile_check
• data_describe, data_query
• run_shell, read_file, write_file, edit_file, append_file
• browse_url, browser_use, send_file, memory_search
• skills_list, skills_activate, skills_read_file

Bundled skills currently shipped in src/researchclaw/agents/skills/ include:

• arxiv
• browser_visible
• citation_network
• cron
• dingtalk_channel
• docx
• experiment_tracker
• figure_generator
• file_reader
• himalaya
• literature_review
• news
• paper_summarizer
• pdf
• pptx
• research_notes
• research_workflows
• xlsx

Workspace model

Runtime data lives under the working directory, while secrets are stored separately:

~/.researchclaw/
├── config.json
├── jobs.json
├── chats.json
├── research/
│   └── state.json
├── sessions/
├── active_skills/
├── customized_skills/
├── papers/
├── references/
├── experiments/
├── memory/
├── md_files/
├── custom_channels/
└── researchclaw.log

~/.researchclaw.secret/
├── envs.json
└── providers.json

Provider credentials and persisted environment variables are intentionally kept out of the working directory.

Development

Backend checks:

pip install -e ".[dev]"
PYTHONPATH=src pytest -q

Console build:

npm --prefix console run build

Website build:

corepack pnpm --dir website run build

Repo-wide helper:

scripts/check-ci.sh --skip-install

Docs

Main documentation files in this repository:

• Quick start
• Deployment
• Console
• Channels
• Skills
• MCP
• Memory
• Config and working dir
• CLI
• FAQ
• Roadmap

Status

The current codebase is best described as:

• already strong on runtime infrastructure, control plane, channels, and provider/skill compatibility
• already usable for persistent research projects, workflow execution, experiment tracking, claim/evidence linking, and blocker handling
• still incomplete as a full autonomous research platform: evidence-matrix quality, rigorous validators, deeper execution backends, and submission packaging remain ahead