Local-first AI agent for UX Research.
ReClaw is an open-source research assistant that runs entirely on your machine. It helps UX researchers organize, analyze, and synthesize research findings using local LLMs — no data ever leaves your computer.
Think OpenClaw meets Google NotebookLM meets Dovetail — but running on your hardware, your models, your data.
Talk to the agent, drop files, trigger skills with natural language.
Every insight traces back: Recommendations > Insights > Facts > Nuggets > Sources.
Create tasks, the agent picks them up and runs the appropriate skill.
Browse, search, edit, and create skills. Track how skills improve over time.
Edit company, project, guardrails, and task context. The agent reads these before every response.
Auto-detects your hardware and recommends the best model. Supports LM Studio and Ollama.
- 42 UXR skills — qualitative and quantitative methods across the full Double Diamond
- Atomic Research — every insight traces back: Recommendations > Insights > Facts > Nuggets > Sources
- RAG on local files — ask questions about your research data with retrieval-augmented generation
- Self-evolving skills — ReClaw analyzes skill performance and proposes improvements automatically
- Chat — conversational interface with skill execution ("analyze my interviews")
- Findings — organized by Double Diamond phase with evidence chain drill-down
- Tasks — Kanban board to direct the agent
- Skills — browse all 42 skills, edit prompts, track self-evolution, create custom skills
- Interviews — transcript viewer with nugget extraction and tag filtering
- Metrics — SUS, NPS, task completion dashboards with benchmarks
- Context — editable 6-layer hierarchy (platform > company > product > project > task > agent)
- Search — Cmd+K global search across all findings
- History — version tracking with rollback
- Settings — hardware info, model management, system status
- Data never leaves your machine — everything runs locally
- LM Studio + Ollama support — choose your preferred LLM provider
- Auto-detects hardware — picks the best model & quantization for your RAM/GPU
- Resource governor — won't overwhelm your machine, reserves resources for other apps
- Token budget management — context window guard with automatic history trimming
- Task Executor — picks Kanban tasks, runs skills, stores findings
- DevOps Audit — monitors data integrity, system health
- UI Audit — heuristic evaluation, accessibility checking
- UX Evaluation — holistic platform experience assessment
- User Simulation — end-to-end API testing
- Meta-Orchestrator — coordinates all agents, prevents conflicts
- Cron Scheduler — schedule recurring skill executions with cron expressions
- Multi-Channel Adapters — extensible Slack, Telegram, and custom channel support
git clone https://github.com/henrique-simoes/ReClaw.git
cd ReClaw
# Backend
cd backend && pip install -e ".[dev]" && cd ..
# Frontend
cd frontend && npm install && cd ..
# Configure LLM provider
cp .env.example .env
# Edit .env to set LLM_PROVIDER=lmstudio or LLM_PROVIDER=ollama# Start your LLM provider (LM Studio or Ollama)
lms server start # LM Studio
# OR: ollama serve # Ollama
# Backend
python -m uvicorn app.main:app --port 8000 --app-dir backend
# Frontend
cd frontend && npm run devThen open http://localhost:3000
cp .env.example .env
mkdir -p data/watch data/uploads data/projects data/lance_db
docker compose upUser Interviews, Contextual Inquiry, Diary Studies, Competitive Analysis, Stakeholder Interviews, Survey Design & Analysis, Analytics Review, Desk Research, Field Studies / Ethnography, Accessibility Audit
Affinity Mapping, Persona Creation, Journey Mapping, Empathy Mapping, JTBD Analysis, HMW Statements, User Flow Mapping, Thematic Analysis, Research Synthesis, Prioritization Matrix, Kappa Intercoder Thematic Analysis, Taxonomy Generator
Usability Testing, Heuristic Evaluation, A/B Test Analysis, Card Sorting, Tree Testing, Concept Testing, Cognitive Walkthrough, Design Critique, Prototype Feedback, Workshop Facilitation
SUS/UMUX Scoring, NPS Analysis, Task Analysis, Regression/Impact Analysis, Design System Audit, Handoff Documentation, Repository Curation, Stakeholder Presentations, Research Retros, Longitudinal Tracking
Skills follow the OpenClaw AgentSkills standard — each is a self-contained directory with SKILL.md, references, and scripts.
6-level system that ensures agents never hallucinate or go off-track:
Level 0: Platform ---- ReClaw UXR expertise (built-in)
Level 1: Company ----- Organization, product, culture, terminology
Level 2: Product ----- Features, users, domain knowledge
Level 3: Project ----- Research questions, goals, timeline
Level 4: Task -------- Per-task instructions from Kanban cards
Level 5: Agent ------- Per-agent system prompts and constraints
Each level is user-editable and composes into the agent's working context. Higher levels override lower levels.
Browser (localhost:3000)
| HTTP/WebSocket
Frontend (Next.js + React + Tailwind)
| REST API + SSE Streaming
Backend (FastAPI + SQLAlchemy + LanceDB)
| OpenAI-compatible API
LM Studio / Ollama (Local LLMs)
| Component | Technology | Why |
|---|---|---|
| Frontend | Next.js 14 + React + Tailwind + Zustand | Rich UI, SSR, great DX |
| Backend | FastAPI (async) + SQLAlchemy | Best AI/ML ecosystem, async, fast |
| Vector Store | LanceDB (embedded) | No extra server, low RAM footprint |
| Database | SQLite (via aiosqlite) | Zero config, reliable, local |
| LLM | LM Studio / Ollama | Hardware detection, multi-model, OpenAI-compatible API |
| Embedding | nomic-embed-text | Runs on CPU, tiny footprint |
| Shortcut | Action |
|---|---|
Cmd+K |
Search findings |
Cmd+1 - Cmd+6 |
Switch views (Chat, Findings, Tasks, Interviews, Context, Skills) |
Cmd+. |
Toggle right panel |
? |
Show keyboard shortcuts |
Esc |
Close modal / cancel |
Enter |
Send message / confirm |
# Backend
cd backend && python -m venv venv && source venv/bin/activate
pip install -e ".[dev]"
uvicorn app.main:app --reload --port 8000
# Frontend
cd frontend && npm install && npm run dev
# LM Studio
lms server start
# OR Ollama
ollama serve && ollama pull qwen3:latestSee CONTRIBUTING.md for guidelines.
- Core platform (chat, findings, tasks, skills)
- 42 UXR skills across all Double Diamond phases
- Multi-agent system with orchestrator
- Context hierarchy and resource governor
- LM Studio + Ollama provider support
- Skills management UI with self-evolution tracking
- Token budget management and context window guard
- Cron scheduler for recurring tasks
- Multi-channel adapter framework (Slack, Telegram)
- URL-based routing and deep linking
- Audio playback with transcript sync
- Full Slack / Telegram integration
- Team features (auth, shared knowledge, access control)
- Multi-model consensus validation (Fleiss' Kappa + cosine similarity)
- Distributed compute via relay nodes
- External LLM server support (Ollama, LM Studio, OpenAI-compatible)
- Adaptive validation method learning
- Dynamic swarm orchestration
- Native installers (dmg, exe, AppImage)
- Skill marketplace
ReClaw's multi-model validation and distributed computing are grounded in peer-reviewed research:
| Feature | Reference | Venue |
|---|---|---|
| Mixture-of-Agents ensemble consensus | Wang et al. Together AI (2025) | ICLR 2025 |
| Self-MoA temperature variation validation | Li et al. (2025) | arXiv 2025 |
| LLM-Blender response aggregation | Jiang et al. (2023) | ACL 2023 |
| Multi-Agent Debate iterative refinement | Du et al. (2024) | ICML 2024 |
| LLM-as-Judge evaluation framework | Zheng et al. (2023) | NeurIPS 2023 |
| Petals distributed LLM inference | Borzunov et al. (2023) | ACL + NeurIPS 2023 |
| Hive volunteer computing for ML | - | SoftwareX 2025 |
| BOINC distributed computing model | Anderson (2020) | - |
| Multi-LLM Thematic Analysis | Jain et al. (2025) | arXiv 2025 |
| Fleiss' Kappa inter-rater reliability | Fleiss (1971) | Psychological Bulletin |
| Atomic Research methodology | Pidcock (2018) | - |
- Consensus Engine: Implements Fleiss' Kappa for categorical agreement and cosine similarity for semantic agreement across multiple model responses. Tiered confidence thresholds by finding type (nuggets κ≥0.70, facts κ≥0.65, insights κ≥0.55, recommendations κ≥0.50).
- Validation Patterns: Five strategies — dual-run, adversarial review, full ensemble, Self-MoA (temperature variation), and debate rounds.
- Adaptive Learning: Tracks which validation method works best per project/skill/agent using weighted scoring with exponential decay recency bias (30-day half-life).
- Distributed Compute: Relay daemon enables team members to donate LLM compute via outbound WebSocket connections (NAT-friendly, no inbound ports). Priority queue ensures user interactions take precedence over background work.
MIT — see LICENSE.
Built with ReClaw by the ReClaw community.