Research Claw
A self-hosted AI research assistant that manages your papers, searches literature, tracks deadlines, and answers on the chat channels you already use.
- Source repo
- nanoAgentTeam/research-claw
- Stars
- ★ 138
- Last updated
- 5mo ago
- License
- MIT
- Primary language
- Python
- FA score
- 58/100 · Major gaps
At a glance
- Works with
- Universal · cross-platformOpenAI API · Claude API
- You'll need
- Typical use
- A solo researcher who wants one conversational entry point for literature review, drafting, and LaTeX compilation instead of juggling an editor, Overleaf, and a search engine.
- Main limitation
- No native Windows support: the project depends on POSIX signal handling and process management, so Windows users must install and configure WSL2 and mind cross-filesystem performance.
- Maintenance
- Commits in the last 6 monthslast commit 2026-06-08
- Source review
- 58/100 · Major gaps
What does this agent do, and when should you use it?
Research Claw (repository nanoAgentTeam/research-claw) is a personal AI research assistant you run on your own machine, MIT-licensed and built on Python 3.11+. It runs in two modes: a CLI for direct terminal interaction, and a Gateway that launches a Web UI (default port 18790) while connecting IM channels including Feishu, Telegram, QQ, and DingTalk — no public IP required. Core components include the AgentLoop main agent, CommandRouter, ContextManager, a ToolRegistry with 40+ tools, and sub-agents that run in isolated sandboxes. It manages LaTeX projects with bidirectional Overleaf sync, and every AI edit is auto-committed to Git for instant rollback. Literature search integrates arXiv, PubMed, and OpenAlex, while radar automation jobs (daily scans, weekly digests, deadline tracking) run on a schedule via APScheduler and push results to your configured channels. On the model side it accepts any OpenAI-compatible API (GPT, DeepSeek, Qwen, Claude, etc.), hot-swappable from the Web UI.
Through chat, Research Claw reads and rewrites .tex / .bib files and compiles LaTeX in one command with automatic error diagnosis. /sync pull and /sync push provide bidirectional Overleaf sync; /git enters an interactive Git mode for history, diff, and rollback, with every AI edit auto-committed. /task decomposes a complex goal into a five-phase pipeline (UNDERSTAND → PROPOSE → PLAN → EXECUTE → FINALIZE), builds a task DAG, and runs sub-agents in parallel inside isolated overlay directories whose outputs are merged before touching the project core. Literature tools call arXiv, PubMed, and OpenAlex and can read full-text PDFs. The Gateway starts APScheduler to fire radar jobs such as Daily Scan, Deadline Watch, and Weekly Digest on cron schedules, pushing results to Telegram, Feishu, DingTalk, Email, or any Apprise channel. Project-level memory persists across sessions, with automated context summarization within token limits. Built-in venue skills cover NeurIPS, ICML, ICLR, AAAI, ACL, CVPR and more, extendable via custom folders under config/.skills/.
- A solo researcher who wants one conversational entry point for literature review, drafting, and LaTeX compilation instead of juggling an editor, Overleaf, and a search engine.
- An author collaborating with an advisor on Overleaf who wants AI edits pushed up and collaborators' changes pulled back automatically.
- A research group tracking multiple conference deadlines and new papers in their field, using Daily Scan, Deadline Watch, and Weekly Digest push notifications.
- A user facing a multi-step writing goal (e.g., a survey) who uses /task to have sub-agents work in parallel — a researcher searching literature, a writer drafting sections — then merged and committed.
- Someone already on Feishu, Telegram, or DingTalk who wants the assistant wired into their existing IM so they can ask questions and trigger tasks from their phone.
- A user who wants all code, papers, and session data kept on their own machine, running self-hosted with no vendor cloud dependency.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
pip install -r requirements.txt
- Config files
settings.json- Runtime
- Python 3.11+
- Install notes cover
- Windows · Linux · WSL
How do you install or deploy this agent?
Requires Python 3.11+, Git, and a LaTeX toolchain. Windows is not supported natively — the project relies on POSIX features, so use WSL2. Keep the code on the WSL Linux filesystem (e.g., ~/) rather than /mnt/c/ for near-native IO performance.
git clone https://github.com/nanoAgentTeam/research-claw.git
cd research-claw
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtFor bidirectional Overleaf sync, run the login command (choose the Overleaf instance, which needs overleaf-sync, or CSTCloud); the generated .olauth is auto-detected:
python cli/main.py loginThe optional browser_use tool requires Chromium:
sudo apt install -y chromium-browserFirst configuration happens in the Web UI: add your LLM provider (API key, model name, base URL — any OpenAI-compatible API works), and optionally IM bot credentials and push subscriptions. All settings are stored in settings.json.
How do you use this agent?
Option A: interact directly in the terminal:
python cli/main.py agentOption B: start the Gateway for the Web UI and IM channels:
python cli/main.py gateway --port 18790Then open http://localhost:18790/ui in your browser to configure providers and channels (under WSL2, ports are forwarded automatically so Windows browsers can connect). Key commands: /list projects, /switch <name>, /compile LaTeX, /sync pull and /sync push for Overleaf, /git for interactive version control, /task <goal> to launch the multi-agent pipeline (review the plan, type /start to execute, /done to exit), /resume for failed tasks, /reset to clear a session. A typical flow: ask to create a project and link Overleaf, ask the bot to research recent papers and draft an introduction, then "/sync push" to upload changes.
What are this agent's strengths and limitations?
- Self-hosted and model-agnostic: any OpenAI-compatible API (GPT, DeepSeek, Qwen, Claude, etc.) works, hot-swappable from the Web UI, with your data staying on your machine.
- Bidirectional Overleaf sync plus automatic Git commits: every AI edit can be pushed and rolled back in seconds, fitting real paper-collaboration workflows.
- Multi-agent task mode decomposes goals into a DAG executed in parallel, with sub-agents in isolated sandboxes whose outputs are merged before touching the project core.
- Built-in research radar automation: Daily Scan, Deadline Watch, Conference Track, and Weekly Digest run on cron schedules and push to your existing IM channels.
- No native Windows support: the project depends on POSIX signal handling and process management, so Windows users must install and configure WSL2 and mind cross-filesystem performance.
- Practical use depends on paid external services: you supply your own LLM API keys, and web/academic search tools may require additional API keys.
- The installation chain is long: Python venv, LaTeX toolchain, optional overleaf-sync, and Chromium all need manual setup — not a one-command install.
- Compilation quality depends on your local LaTeX setup, and the docs offer no containerized (Docker) deployment path, leaving environment-drift troubleshooting to you.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Research Claw This agent | 58 · Major gaps | — | ★ 138 | 5mo ago | Python | OpenAI API · Claude API |
| PocketPaw | 44 · Major gaps | Desktop appFree + model costs | ★ 935 | 3d ago | Python | Codex · OpenAI API · Claude API |
| Sapphire Blue AI | 69 · Some gaps | Web appFree + model costs | ★ 336 | 14d ago | Python | OpenAI API · Claude API |
| NextClaw | 61 · Some gaps | Desktop appFree + model costs | ★ 264 | 1d ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Skills whitelist tools via allowed-tools; sub-agents run in isolated overlay directories merged only after approval; CLAUDE.md mandates path-traversal-safe file APIs — good least-privilege posture. Task mode requires /start approval. Deductions: scheduled radar jobs auto-create and proactively push to Telegram/Feishu/Email, and AI edits can be pushed to Overleaf — user-confirmation mechanisms for these automated external effects are not shown; API keys and IM credentials stored in plaintext settings.json with no encryption mentioned; requirements.txt is entirely lower-bound unpinned with browser-use and pyoverleaf completely unpinned and no dependency audit; rollback is well evidenced via auto Git commit per edit plus interactive /git rollback. LICENSE copyright reads 'ContextBot Contributors', inconsistent with nanoAgentTeam/research-claw, weakening attribution.
CLI gives clear errors for missing dependencies and failed logins with install hints; IM handlers log and gracefully degrade on send failure. Deductions: naming self-inconsistency (package open-research-claw, LICENSE 'ContextBot', app open_research_claw vs brand Research Claw), and the sampled QQ/DingTalk channel code is a ping/echo demo inconsistent with README's full-channel claims. Optional dependency detection (overleaf-sync, PySide6) is handled adequately.
Audience (self-hosted researchers) and scenarios are clearly described; bilingual guides included. Environment fit is excellent: full WSL2 walkthrough, /mnt/c performance warning, port forwarding, optional Chromium, explicit Windows-non-native statement. Skill triggers are concrete but the 'user mentions X' activation pattern risks false triggers on multi-channel chat; tool availability is scoped per workspace (Lobby vs Project), giving decent capability boundaries.
Install documentation is thorough (full WSL path) — a clear strength; architecture diagram, config reference, command table, bilingual guides all present; MIT license exists and matches README (copyright-holder name inconsistent, minor deduction). Weaknesses: version 0.1.0 with no CHANGELOG; names differ across repo/package/product/LICENSE holder; known limitations limited to the Windows statement with nothing on token cost, API-key security, or automated-push misfires; maintenance responsibility is only a contributing section with no SECURITY/maintainer/release process; no FAQ.
Output usability is reasonable: streamed task progress, /git diff/rollback, compile-error auto-diagnosis. Marginal value for the target user is evident (Overleaf + literature search + IM channels integrated). Cost/benefit is undiscussed: 40+ tools, docker, browser-use, MarkItDown[all] heavy deps and long LLM sessions carry real cost with zero user guidance; sampled test timeout_sec=3600 hints at very long runs.
README links two public demo papers, partially corroborating capability claims; tests/cases specify expected tool-call sequences and file artifacts, reviewable. But claims like '40+ tools' and 'real user demo' lack in-sample evidence, and channel sample code contradicts marketing copy — weak cross-source corroboration. The conference-research skill explicitly requires >=3 papers as evidence, separates Main/Findings/Workshop, and forbids metadata-only assertions: good fact/inference separation discipline.
- API keys and IM credentials are stored in plaintext settings.json; restrict file permissions on self-hosted deployments and never commit them.
- requirements.txt dependencies are almost entirely unpinned (browser-use, pyoverleaf have no version at all); pin and audit before deploying.
- Scheduled radar jobs run automatically and push to external channels; Overleaf push modifies remote projects — verify push subscriptions and disable auto-created radar jobs if unwanted.
- Sampled QQ/DingTalk channel code is only a ping/echo demo, contradicting README channel claims; verify actual channel integration before production use.
- Heavy dependencies (docker, browser-use, MarkItDown[all]) and long LLM sessions are undocumented cost drivers; monitor token and resource usage.
FAQ
Is Research Claw open source?
Is Research Claw safe to use?
Is Research Claw still maintained?
Who makes Research Claw?
Thanks — we use this to improve the page.