Openclaw Research Tool

Search the web using LLMs via OpenRouter. Use for current web data, API docs, market research, news, fact-checking, or any question that benefits from live i...

Security Scan

The skill's requirements and runtime instructions are coherent with a CLI research tool that uses OpenRouter; it asks only for an OpenRouter API key and the research-tool binary, which match its stated purpose.

Purpose & Capability

Name/description (web research via OpenRouter) align with required items: the skill needs the research-tool CLI binary and an OPENROUTER_API_KEY. The README/metadata references a GitHub repo, which is consistent with a third-party CLI wrapper.

Instruction Scope

SKILL.md contains only instructions to run the research-tool CLI (sessions_spawn/exec) and how to set flags; it does not instruct reading unrelated local files or secrets. It does reference optional env vars (RESEARCH_EFFORT, RESEARCH_MODEL) that are not listed in requires.env — these are configuration knobs, not extra secrets. It also warns not to set exec timeout (operational guidance) which is unusual but not a security issue by itself.

Install Mechanism

The skill is instruction-only (no install spec). SKILL.md suggests installing via `cargo install openclaw-search-tool`, which is a reasonable, traceable install mechanism for a Rust-based CLI. The registry did not include an automated install step; that's low risk but means users should verify the binary source themselves before installing.

Credentials

Only OPENROUTER_API_KEY is required (declared as primary credential), which is proportional to a tool that talks to OpenRouter. The docs mention additional optional env vars (RESEARCH_MODEL, RESEARCH_EFFORT) not declared in requires.env — these are non-sensitive configuration items. Users should remember the API key grants access/billing on OpenRouter and therefore is sensitive.

Persistence & Privilege

The skill does not request always:true or any system-wide config paths and does not attempt to persist beyond normal CLI usage. disable-model-invocation is false (normal) — the skill can be invoked by the agent but has no elevated installation privileges.

Assessment

This skill appears to do what it says: run a CLI that queries OpenRouter and returns citation-backed answers. Before installing: (1) verify the origin of the research-tool binary (inspect the GitHub repo or crate) because the CLI will send your queries to OpenRouter; (2) treat OPENROUTER_API_KEY as a secret (it enables access and billing on your OpenRouter account); (3) avoid sending sensitive personal or proprietary data through the tool unless you trust OpenRouter's handling and your account settings; (4) follow the author's recommendation to run it in a sub-agent so your main session doesn't block; and (5) if you want stronger assurance, review the CLI source code or build from source rather than running a prebuilt binary.

Web search for OpenClaw agents, powered by OpenRouter. Ask questions in natural language, get accurate answers with cited sources. Defaults to GPT-5.2 which excels at documentation lookups and citation-heavy research.

Note: Even low-effort queries may take 1 minute or more to complete. High/xhigh reasoning can take 10+ minutes depending on complexity. This is normal — the model is searching the web, reading pages, and synthesizing an answer.

Recommended: Run research-tool in a sub-agent so your main session stays responsive:

sessions_spawn task:"research-tool 'your query here'"

⚠️ Never set a timeout on exec when running research-tool. Queries routinely take 1-10+ minutes. Use yieldMs to background it, then poll — but do NOT set timeout or the process will be killed mid-search.

The :online model suffix gives any model live web access — it searches the web, reads pages, cites URLs, and synthesizes an answer.

Install

cargo install openclaw-search-tool

Requires OPENROUTER_API_KEY env var. Get a key at https://openrouter.ai/keys

Quick start

research-tool "What are the x.com API rate limits?"
research-tool "How do I set reasoning effort parameters on OpenRouter?"

From an OpenClaw agent

# Best: run in a sub-agent (main session stays responsive)
sessions_spawn task:"research-tool 'your query here'"

# Or via exec — NEVER set timeout, use yieldMs to background:
exec command:"research-tool 'your query'" yieldMs:5000
# then poll the session until complete

Flags

--effort, -e (default: low)

Controls how much the model reasons before answering. Higher effort means better analysis but slower and more tokens.

research-tool --effort low "What year was Rust 1.0 released?"
research-tool --effort medium "Explain how OpenRouter routes requests to different model providers"
research-tool --effort high "Compare tradeoffs between Opus 4.6 and gpt-5.3-codex for programming"
research-tool --effort xhigh "Deep analysis of React Server Components vs traditional SSR approaches"

Level	Speed	When to use
low	~1-3 min	Quick fact lookups, simple questions
medium	~2-5 min	Standard research, moderate analysis
high	~3-10 min	Deep analysis with careful reasoning
xhigh	~5-20+ min	Maximum reasoning, complex multi-source synthesis

Can also be set via env var RESEARCH_EFFORT.

--model, -m (default: openai/gpt-5.2:online)

Which model to use. Defaults to GPT-5.2 with the :online suffix because it excels at questions where citations and accurate documentation lookups matter. The :online suffix enables live web search and works with any model on OpenRouter.

# Default: GPT-5.2 with web search (great for docs and cited answers)
research-tool "current weather in San Francisco"

# Claude with web search
research-tool -m "anthropic/claude-sonnet-4-20250514:online" "Summarize recent changes to the OpenAI API"

# GPT-5.2 without web search (training data only)
research-tool -m "openai/gpt-5.2" "Explain the React Server Components architecture"

# Any OpenRouter model
research-tool -m "google/gemini-2.5-pro:online" "Compare React vs Svelte in 2026"

Can also be set via env var RESEARCH_MODEL.

--system, -s

Override the system prompt to give the model a specific persona or instructions.

research-tool -s "You are a senior infrastructure engineer" "Best practices for zero-downtime Kubernetes deployments"
research-tool -s "You are a Rust systems programmer" "Best async patterns for WebSocket servers"

--stdin

Read the query from stdin. Useful for long or multiline queries.

echo "Explain the OpenRouter model routing architecture" | research-tool --stdin
cat detailed-prompt.txt | research-tool --stdin

--max-tokens (default: 12800)

Maximum tokens in the response.

--timeout (optional, no default)

No timeout by default — queries run until the model finishes. Set this only if you need a hard upper bound (e.g. --timeout 300).

Output format

• stdout: Response text only (markdown with citations) — pipe-friendly
• stderr: Progress status, reasoning traces, and token usage

🔍 Researching with openai/gpt-5.2:online (effort: high)...
✅ Connected — waiting for response...

[response text on stdout]

📊 Tokens: 4470 prompt + 184 completion = 4654 total | ⏱ 5s

Status indicators

• 🔍 Researching... — request sent to OpenRouter
• ✅ Connected — waiting for response... — server accepted the request, model is searching/thinking
• ⏳ 15s... ⏳ 30s... — elapsed time ticks (only in interactive terminals, not in agent exec)
• ❌ Connection to OpenRouter failed — couldn't reach OpenRouter (network issue)
• ❌ Connection to OpenRouter lost — connection dropped while waiting. Retry?

Tips for better results

• Write in natural language. "What are the best practices for Rust error handling and when should you use anyhow vs thiserror?" works better than keyword-style queries.
• Provide maximum context. The model starts from zero. Include background, what you already know, and all related sub-questions. Detailed prompts massively outperform vague ones.
• Use effort levels appropriately. low for quick facts, high for real research, xhigh only for complex multi-source analysis.
• Use -s for domain expertise. A specific persona produces noticeably better domain-specific answers.

Cost

~$0.01–0.05 per query. Token usage is printed to stderr after each query.