@sinlair/reagent

🧭 ReAgent

Turn ongoing research into a reusable workflow, not a one-off conversation.

🧠 Local-first research workspace for paper discovery, evidence review, research notes, and delivery.

🇺🇸 English  |  🇨🇳 简体中文

🛣️ Roadmap  •  🤝 Contributing  •  🛠️ Operations

⚡ At A Glance

• 🧭 Plan research around reusable templates instead of ad-hoc prompts.
• 🔬 Run tasks, inspect evidence, and review outputs in one workspace.
• 🧠 Build a file-backed knowledge base with indexing, recall, and compaction.
• 📦 Ship reusable outputs such as reports, slides, and module files.

📌 What ReAgent Is

ReAgent is not a chatbot with a few research buttons attached. It is a local research workspace for people who need to revisit topics, inspect evidence, preserve context, and ship useful outputs.

It is designed for workflows such as:

• 🧭 Defining reusable research templates with goals, background, and success criteria.
• 🔎 Running repeated discovery instead of restarting the same search every week.
• 📚 Reviewing papers, repo findings, feedback, and notes in one place.
• 🧠 Preserving working context across tasks instead of losing it inside chat history.
• 🧾 Turning research into reports, briefings, slides, and reusable artifacts.

✨ Highlights

Icon	Highlight	Why it matters
🧭	Research templates	Save a reusable setup for a topic before spending time or compute.
🔬	Research workspace	Run tasks, inspect progress, read reports, and manage outputs in one page.
🧠	Research notes	Keep durable notes outside chat so context survives across sessions.
🗺️	Research map	Trace how topics, evidence, reports, and files connect.
📦	Reusable outputs	Produce topic reports, slides, module files, and other shareable artifacts.
🔌	Multi-surface access	Use the web UI, WeChat, and OpenClaw plugin without splitting the workflow.
⚙️	Local runtime control	Inspect channels, sessions, models, skills, logs, and deployment posture locally.

🧠 How The Knowledge Base Is Built

ReAgent's knowledge base is file-backed first, then indexed for recall.

• 📝 Long-term notes are stored in MEMORY.md.
• 🗓️ Daily notes are stored in memory/YYYY-MM-DD.md.
• 🗂️ Every saved note is indexed into memory-index.json with title, snippet, source, confidence, tags, entity IDs, and timestamps.
• 🔎 Search works on real workspace files, while recall can merge indexed workspace memory with durable artifacts such as research templates, topic reports, and presentations.
• ♻️ Direction reports can flush high-signal summaries back into memory, so later tasks can recall them as context.
• 🧹 Memory policy lives in memory-policy.json, and older daily notes can be compacted into long-term summaries.
• 🧾 Every compaction is recorded in memory-compactions.json, so the history of what was folded and why remains visible.

The practical result is a local knowledge base with four layers working together:

• 📄 Raw Markdown notes for direct inspection.
• 🗃️ Indexed metadata for fast recall.
• 📚 Durable research artifacts that can also be recalled.
• 🧹 Compaction history and policy so memory does not become unmanageable over time.

🖥️ Product Surfaces

The current web UI is organized around a small set of working surfaces:

• 🏠 Workbench: current progress, latest results, and common next actions.
• 💬 Main Workspace: chat, quick research actions, recent tasks, and current workspace context.
• 📈 Status: activity, latest outputs, and system health.
• 🔬 Research Workspace: templates, task lists, reports, schedules, topic reports, slides, and module files.
• 🗺️ Research Map: how topics, evidence, reports, and files connect.
• 🧠 Research Notes: search, save, and reopen file-backed notes.
• 🔌 Channels: WeChat login, connection status, lifecycle changes, and channel events.
• 🤖 Agent Settings and Skills: choose the role, model, and tool access used by chat.

Single CLI Surface

The browser UI mirrors this CLI-first runtime surface. The Runtime & Delivery and Runtime Logs panels are intended as observability and copy-friendly control panels, not as a separate control plane.

ReAgent is converging on reagent as the only required control surface.

• Use reagent service ... to control the supervised runtime and reagent runtime ... to inspect it.
• Use reagent research ... for tasks, directions, discovery, graph inspection, reports, and artifacts.
• Use reagent channels ..., reagent memory ..., reagent config ..., and reagent plugins ... for the rest of the workspace runtime.
• Treat the web UI as an optional inspection surface that mirrors the CLI runtime surface, not the only place where state can be changed.
• reagent gateway ..., reagent status, reagent health, and reagent logs still work as compatibility aliases.

Common CLI flows:

reagent service run
reagent runtime status
reagent runtime logs --follow
reagent research recent
reagent research tasks
reagent research directions
reagent channels status
reagent memory recall "recent research choices"
reagent config validate

🚀 Quick Start

• 📄 Create a local env file from .env.example.
• 📦 Install dependencies.
• 🗃️ Push the Prisma schema to the default SQLite database.
• ▶️ Start the dev server.

npm install
npm run db:push
npm run dev

🌐 Open http://127.0.0.1:3000/.

🧪 Minimal local setup:

LLM_PROVIDER=fallback
WECHAT_PROVIDER=mock

🪟 If PowerShell blocks npm, use npm.cmd instead.

🧰 Global CLI install:

npm install -g @sinlair/reagent
reagent init
reagent service run

🏷️ The published package name is @sinlair/reagent, while the installed command is reagent.

From there, stay inside the CLI:

reagent runtime status
reagent runtime logs --follow
reagent research recent
reagent research graph report
reagent channels login
reagent plugins list

🔁 Always-on gateway lifecycle:

reagent service install
reagent service status
reagent service restart
reagent service stop

🔄 Run Modes

The root app can run as a foreground dev process or as an always-on background service.

Icon	Mode	Commands	Notes
🧪	Development	npm run dev	Local development with live reload
♾️	PM2	npm run pm2:start / npm run pm2:restart / npm run pm2:logs	Keep the app running in the background
🪟	Windows Service	npm run service:install / npm run service:status / npm run service:start / npm run service:stop	Machine-level runtime on Windows

See OPERATIONS.md for deployment and maintenance details.

🗂️ Repository Layout

Icon	Path	Purpose
🧩	./	Root ReAgent app, web UI, API server, and runtime
🧠	packages/reagent-core/	Reusable core research logic
🔌	packages/reagent-openclaw/	Installable OpenClaw plugin package
🧪	package/	In-repo OpenClaw WeChat reference package kept for compatibility work
🖼️	docs/	Product visuals and supporting docs

✅ Development

Validation:

npm run check:all
npm run test

If you only want to validate the publishable packages:

npm run build:packages
npm run check:packages

🔌 OpenClaw Plugin

Install the ReAgent OpenClaw plugin with:

openclaw plugins install @sinlair/reagent-openclaw --yes

Plugin source lives in packages/reagent-openclaw/.

📚 Documentation

• 📘 Chinese README: README.zh-CN.md
• 🧭 Product blueprint: agent.md
• 🛣️ Roadmap: ROADMAP.md
• 🛠️ Operations: OPERATIONS.md
• 🤝 Contributing: CONTRIBUTING.md
• 🔒 Security: SECURITY.md

🙌 Inspiration

ReAgent is shaped by research-agent and research-workspace projects such as:

• 🧪 GPT Researcher
• 🌊 deer-flow
• 🧭 PASA
• 📄 Paper2Agent
• 🏢 enterprise-deep-research
• 🤖 InternAgent
• 🔓 OpenClaw

For a closer comparison, see docs/research-agent-landscape.md.

📄 License

Released under the MIT License.