scientify

Scientify

AI-powered research workflow automation for OpenClaw.

Scientify is an OpenClaw plugin that automates the full academic research pipeline — from literature survey to experiment execution — using LLM-driven sub-agents.

Website: scientify.tech | 中文文档

What It Does

Scientify turns a single research prompt into a complete automated pipeline. Each phase runs as an independent sub-agent — the orchestrator verifies outputs between steps and passes context forward.

Scenario 1 — End-to-End Research Pipeline

"Research scaling laws for classical ML classifiers on Fashion-MNIST"

The research-pipeline orchestrator runs all 6 phases in sequence, spawning a dedicated sub-agent for each:

flowchart LR
    A["Literature\nSurvey"] --> B["Deep\nAnalysis"] --> C["Implementation\nPlan"] --> D["Code\nImplementation"] --> E["Automated\nReview"] --> F["Full\nExperiment"]

What each phase produces

Phase	What Happens	Output File
1. Literature Survey	Search arXiv + OpenAlex, filter, download .tex sources, cluster by direction	survey/report.md
2. Deep Analysis	Extract formulas, map methods to code, build cross-comparison	survey_res.md
3. Implementation Plan	Design 4-part plan — Dataset / Model / Training / Testing	plan_res.md
4. Code Implementation	Write ML code in uv-isolated venv, validate with 2-epoch run	project/run.py
5. Automated Review	Review code → fix issues → rerun → re-review (up to 3 rounds)	iterations/judge_v*.md
6. Full Experiment	Complete training + ablation studies with final analysis	experiment_res.md

Scenario 2 — Idea Generation

"Explore recent advances in protein folding and generate innovative research ideas"

The idea-generation skill surveys the field, then:

• Generates 5 diverse research ideas grounded in real papers
• Scores each on novelty, feasibility, and impact
• Selects the best and produces an enhanced proposal with detailed methodology

[!TIP] Output: ideas/selected_idea.md — a ready-to-develop research proposal.

Scenario 3 — Standalone Literature Survey

"Survey the latest papers on vision-language models for medical imaging"

Run just the survey phase when you need a structured reading list without running the full pipeline:

• Searches arXiv (CS/ML) and OpenAlex (cross-disciplinary, broader coverage)
• Downloads .tex source files; retrieves open-access PDFs via Unpaywall
• Clusters papers by sub-topic and extracts key findings
• Generates a structured survey report

[!TIP] Output: survey/report.md + raw papers in papers/_downloads/

Scenario 4 — Review Paper Drafting

"Write a survey paper based on my project's research outputs"

After completing a research pipeline (or just a literature survey + deep analysis), the write-review-paper skill assembles a draft:

• Synthesizes survey reports, analysis notes, and comparison tables
• Structures the paper with Introduction, Related Work, Methods, and Discussion
• Produces a publication-ready draft in Markdown

[!TIP] Output: a survey/review paper draft based on all accumulated project artifacts.

Advanced Scenarios — Combining OpenClaw Platform Capabilities

As an OpenClaw plugin, Scientify can leverage the platform's MCP servers, browser automation, multi-session concurrency, and more to build powerful composite workflows.

Scenario 5 — Literature Monitoring Bot

"Automatically search for new diffusion model papers every day and push a digest to our Slack channel"

Combine OpenClaw's MCP integration (Slack / Feishu / Email) with scheduled triggers to build automated literature monitoring:

flowchart LR
    A["Scheduled Trigger\n(cron / webhook)"] --> B["arxiv_search\n+ openalex_search"]
    B --> C["LLM Filtering\n+ Summary"]
    C --> D["Push to\nSlack / Feishu / Email"]

• External cron job or OpenClaw webhook triggers a session periodically
• Scientify's arxiv_search + openalex_search fetch the latest papers
• LLM scores and filters by your research interests, generates concise summaries
• MCP tools push the digest to Slack, Feishu, or email

[!NOTE] Requires: A configured MCP server (e.g., slack-mcp, feishu-mcp). OpenClaw supports declaring MCP servers in openclaw.json.

Scenario 6 — Download Paywalled Papers via Browser

"Download these 5 IEEE papers using my university VPN"

Scientify's built-in arxiv_download and unpaywall_download only handle open-access papers. For paywalled content, combine with OpenClaw's browser automation (Playwright MCP):

flowchart LR
    A["Scientify\nprovides paper URLs"] --> B["Playwright MCP\nopens browser"]
    B --> C["Institutional Proxy\nauto-authenticate"]
    C --> D["Navigate to Publisher\ndownload PDF"]

• OpenClaw launches a controlled browser via Playwright MCP server
• The browser accesses publisher sites through your institutional proxy / VPN
• Automatically navigates to the paper page and downloads the PDF to papers/_downloads/
• Works with IEEE, Springer, Elsevier, ACM, and other subscription-based publishers

[!NOTE] Requires: Playwright MCP server configured, and institutional network access to the papers.

Scenario 7 — Multi-Topic Parallel Research

"Research 3 directions simultaneously: LoRA fine-tuning, MoE architectures, KV-Cache optimization"

Leverage OpenClaw's multi-session concurrency (sessions_spawn) to run multiple research pipelines in parallel:

flowchart TD
    O["Main Agent\n(Orchestrator)"] --> A["Sub-session 1\nLoRA Fine-tuning"]
    O --> B["Sub-session 2\nMoE Architectures"]
    O --> C["Sub-session 3\nKV-Cache Optimization"]
    A --> D["Independent project dirs\nisolated from each other"]
    B --> D
    C --> D

• Each sub-topic runs a full pipeline with its own project directory
• The main agent collects results and produces a cross-topic comparative analysis
• Ideal for quickly scouting multiple directions during the topic-selection phase of a survey paper

Scenario 8 — Interactive Paper Reading Assistant

"Walk me through 'Attention Is All You Need' section by section, explain every formula"

Combine OpenClaw's conversational interface with Scientify's paper_browser tool for interactive deep reading:

• paper_browser loads papers page-by-page, avoiding context overflow
• Discuss section by section: LLM explains derivations, compares with related work, highlights contributions
• Follow up on implementation details — LLM uses github_search to find corresponding open-source code
• All analysis notes are saved to notes/paper_{id}.md

Scenario 9 — Paper-to-Reproducible-Experiment

"Reproduce the results from Table 2 of this paper"

End-to-end automation: understand paper → implement code → run experiment → compare results:

flowchart LR
    A["paper_browser\nDeep read paper"] --> B["research-plan\nExtract experiment design"]
    B --> C["research-implement\nWrite code"]
    C --> D["research-experiment\nRun experiment"]
    D --> E["Compare with\npaper's Table 2"]

• paper_browser reads the method and experiment sections in detail
• research-plan extracts experiment config (hyperparameters, datasets, metrics)
• research-implement generates code and validates in a uv-isolated environment
• research-experiment runs the full experiment
• LLM automatically compares your results against the paper's reported numbers

Prerequisites

• Node.js >= 18
• Python 3 + uv (for ML code execution)
• git

Install OpenClaw

# Install OpenClaw globally
pnpm add -g openclaw    # or: npm install -g openclaw

# Run onboarding wizard (configures model provider, API key, workspace)
openclaw onboard

# Start the gateway (runs the WebUI server)
openclaw gateway

After openclaw gateway, the WebUI is available at http://127.0.0.1:18789/ (default port).

Proxy users: If you have http_proxy set, access the WebUI with --noproxy 127.0.0.1 or configure your browser accordingly.

Install Scientify

From npm (recommended)

openclaw plugins install scientify

The plugin installs to ~/.openclaw/extensions/scientify/ and is automatically enabled.

From source (development)

git clone https://github.com/user/scientify.git
cd scientify && pnpm install && pnpm build

# Link as dev plugin
openclaw plugins install -l ./

Verify installation

openclaw plugins list
# Should show: scientify (enabled)

After installation, restart the gateway to load the plugin:

# Stop the running gateway (Ctrl+C), then:
openclaw gateway

Usage via WebUI

1. Open the WebUI

Navigate to http://127.0.0.1:18789/ in your browser.

2. Start a research task

Type a research prompt in the chat. Scientify skills are auto-matched by the LLM:

Research "transformer efficiency" and generate some innovative ideas

Or invoke a specific skill directly with a slash command:

/research-pipeline
/research-collect
/idea-generation

3. Monitor sub-agent progress

When the orchestrator spawns sub-agents, you'll see:

• Spawn notification — "Phase 1: Literature Survey started"
• Completion announcement — automatic message when the sub-agent finishes
• Progress updates — the orchestrator verifies outputs and advances to the next phase

You can also check status anytime with:

/research-status

4. Manage projects

/projects              # List all projects
/project-switch <id>   # Switch to a different project
/papers                # List downloaded papers
/ideas                 # List generated ideas

Skills

Pipeline Skills (LLM-powered)

Skill	Slash Command	Description
research-pipeline	/research-pipeline	Orchestrator. Spawns sub-agents for each phase, verifies outputs between steps.
research-collect	/research-collect	Search arXiv → filter → download .tex sources → cluster → generate survey report.
research-survey	/research-survey	Deep analysis of papers: extract formulas, map to code, produce method comparison table.
research-plan	/research-plan	Create 4-part implementation plan (Dataset/Model/Training/Testing) from survey results.
research-implement	/research-implement	Implement ML code from plan, run 2-epoch validation with uv venv isolation.
research-review	/research-review	Review implementation. Iterates fix → rerun → review up to 3 times.
research-experiment	/research-experiment	Full training + ablation experiments. Requires review PASS.
idea-generation	/idea-generation	Generate 5 innovative research ideas from a topic, select and enhance the best one.

Standalone Skills

Skill	Description
write-review-paper	Draft a review/survey paper from project research outputs.
research-subscription	Create/list/remove scheduled Scientify jobs via scientify_cron_job (research digests or plain reminders).

Tools (available to LLM)

Tool	Description
arxiv_search	Search arXiv papers. Returns metadata (title, authors, abstract, ID). Does not download files. Supports sorting by relevance/date and date filtering.
arxiv_download	Batch download papers by arXiv ID. Prefers .tex source files (PDF fallback). Requires absolute output directory path.
openalex_search	Search cross-disciplinary academic papers via OpenAlex API. Returns DOI, authors, citation count, OA status. Broader coverage than arXiv.
openreview_lookup	Lookup OpenReview evidence by title/ID/forum. Returns decision (if available), review rating/confidence aggregates, and concise review summaries for venue-risk analysis.
unpaywall_download	Download open access PDFs by DOI via Unpaywall API. Non-OA papers are silently skipped (no failure).
github_search	Search GitHub repositories. Returns repo name, description, stars, URL. Supports language filtering and sorting.
paper_browser	Paginated browsing of large paper files (.tex/.md) to avoid loading thousands of lines into context. Returns specified line range with navigation info.
scientify_cron_job	Manage scheduled Scientify jobs from the model (upsert/list/remove). Main fields: action, scope, schedule, topic, project, message, max_papers, recency_days, candidate_pool, score_weights, sources, channel, to, no_deliver, run_now, job_id. run_now=true also returns a status_json snapshot for research tasks.
scientify_literature_state	Persistent incremental state for subscriptions: prepare dedupe context (+ memory hints), record pushed papers + project knowledge_state artifacts (including paper_notes deep-reading fields and full-text cleanup run logs), feedback lightweight preference memory, and status inspection with traceable logs.

Commands (direct, no LLM)

Command	Description
/research-status	Show workspace status and active project
/papers	List downloaded papers with metadata
/ideas	List generated ideas
/projects	List all projects
/project-switch <id>	Switch active project
/project-delete <id>	Delete a project
/research-subscribe ...	Create/update scheduled Scientify jobs (supports daily, weekly, every, at, cron; options: --channel, --to, --topic, --project, --message, --max-papers, --recency-days, --candidate-pool, --score-weights, --sources, --no-deliver)
/research-subscriptions	Show your scheduled Scientify jobs
/research-unsubscribe [job-id]	Remove your scheduled Scientify jobs (or a specific job)
/metabolism-status	Compatibility alias view over project knowledge_state summary

/research-subscribe examples:

• /research-subscribe daily 09:00 Asia/Shanghai (auto-deliver to current chat sender/channel when possible)
• /research-subscribe every 2h --channel feishu --to ou_xxx
• /research-subscribe at 2m --channel feishu --to ou_xxx
• /research-subscribe weekly mon 09:30 --channel telegram --to 123456789
• /research-subscribe at 2m --channel webui (webui/tui are aliases of last)
• /research-subscribe daily 08:00 --topic "LLM alignment"
• /research-subscribe daily 08:00 --topic "LLM alignment" --project llm-alignment
• /research-subscribe daily 08:00 --topic "LLM alignment" --max-papers 5 --recency-days 30 --sources arxiv,openalex
• /research-subscribe daily 08:00 --topic "LLM alignment" --candidate-pool 12 --score-weights relevance:45,novelty:20,authority:25,actionability:10
• /research-subscribe at 1m --message "Time to drink coffee."
• /research-subscribe daily 09:00 --no-deliver (background only, no push)

Behavior notes:

• Scoped upsert: per sender/channel scope, creating a new subscription replaces the previous one in that scope.
• Delivery aliases: --channel webui and --channel tui map to last; they do not require --to.
• Reminder-safe fallback: if topic looks like a plain reminder (for example "remind me to sleep"), Scientify auto-routes it as a reminder message instead of literature pipeline.
• One-shot topic (at ... --topic ...) uses focused retrieval of representative papers; recurring schedules (daily/weekly/every/cron) use incremental tracking mode.
• Recurring incremental mode uses candidate-pool ranking before Top-K selection; if no unseen paper is found, it runs one representative fallback pass before returning empty.
• Default max_papers is 5 unless overridden by --max-papers.
• Built-in quality gates use soft mode by default: non-fatal quality gaps are stored as warnings (quality_gate.severity=warn) while the run remains ok; only fatal issues are downgraded to degraded_quality.
• Default strict recall behavior uses tiered retrieval (topic-related + broader method + adjacent cross-domain) to avoid sparse core selection on broad topics.
• Each run executes at most one immediate reflection follow-up and writes trace/results back to knowledge_state for auditability.
• Hypothesis-facing response policy is gate-driven:
  • if knowledge_state_summary.hypothesis_gate.accepted == 0, Scientify returns factual cycle status only (no speculative roadmap/deep-dive suggestions)
  • if accepted > 0, Scientify includes hypothesis details in the current message by default
  • when channel/runtime clearly supports multi-send, it may optionally split into two consecutive pushes (brief alert + detailed hypothesis)
• Lightweight preference memory is stored backend-only (keyword/source affinities) and used to rerank future pushes quietly.
• Incremental dedupe + memory state is persisted under ~/.openclaw/workspace/scientify/ (literature-state.json, literature-push-log.jsonl).
• Project-level research traceability is persisted under ~/.openclaw/workspace/projects/{project-id}/knowledge_state/.
• Full-text-first scheduled runs store per-paper deep-reading notes (domain/subdomains/cross-domain links/research goal/approach/design/contributions/practical insights/must-understand points/limitations/evidence anchors) under knowledge_state/paper_notes/.
• Full-text files should be downloaded to a temporary directory and cleaned after each run; cleanup result is tracked in knowledge_state run logs.
• Storage: subscription jobs are stored in OpenClaw cron storage; knowledge artifacts are stored in project workspace files.
• Global inspect: openclaw cron list --all --json
• Compatibility aliases: openclaw research init/list/status/delete and /metabolism-status are retained for migration, but internally use the same unified projects/*/knowledge_state architecture.

Workspace Structure

All research data is organized under ~/.openclaw/workspace/projects/:

projects/
├── .active                        # Current project ID
├── scaling-law-fashion-mnist/     # Example project
│   ├── project.json               # Metadata
│   ├── task.json                  # Task definition
│   ├── papers/
│   │   ├── _meta/                 # Paper metadata (*.json)
│   │   └── _downloads/            # Raw .tex/.pdf files
│   ├── survey/
│   │   └── report.md              # Literature survey report
│   ├── notes/                     # Per-paper deep analysis
│   │   └── paper_{arxiv_id}.md
│   ├── survey_res.md              # Method comparison table
│   ├── plan_res.md                # Implementation plan
│   ├── project/                   # ML code
│   │   ├── run.py
│   │   └── requirements.txt
│   ├── ml_res.md                  # Implementation results
│   ├── iterations/                # Review iterations
│   │   └── judge_v*.md
│   ├── experiment_res.md          # Final experiment results
│   ├── ideas/                     # Generated ideas
│       ├── idea_*.md
│       └── selected_idea.md
│   └── knowledge_state/           # Scheduled research state artifacts
│       ├── knowledge/
│       ├── paper_notes/           # Per-paper deep-reading records
│       ├── daily_changes/
│       ├── hypotheses/
│       ├── logs/
│       ├── state.json
│       └── events.jsonl
└── another-project/

Configuration

Plugin settings in ~/.openclaw/openclaw.json:

{
  "plugins": {
    "entries": {
      "scientify": {
        "enabled": true
      }
    }
  }
}

Plugin management

openclaw plugins list               # List installed plugins
openclaw plugins enable scientify    # Enable
openclaw plugins disable scientify   # Disable
openclaw plugins update scientify    # Update to latest
openclaw plugins doctor              # Diagnose issues

Known Limitations

• Sub-agent timeout: Each sub-agent has a 30-minute timeout (runTimeoutSeconds: 1800). Complex literature surveys with many papers may need longer.
• GPU/Sandbox: Code execution runs on host by default. OpenClaw sandbox does not support GPU passthrough yet.
• Model dependency: Research quality depends heavily on the LLM model used. Claude Opus 4.5+ or GPT-5+ recommended.

Development

git clone https://github.com/user/scientify.git
cd scientify
pnpm install
pnpm build          # Build TypeScript
pnpm dev            # Watch mode

# Link to OpenClaw for testing
openclaw plugins install -l ./

See CLAUDE.md for version update SOP and contribution guide.

License

MIT

Author

tsingyuai