odd-bot-research

A deep research stack for OpenClaw agents. Seven MCP servers, six audited skills, and one orchestrator that plans, fans out, cross-verifies, and writes citation-backed reports.

GitHub repo Agent instructions (plain text) Troubleshooting

Setting this up with an agent? Paste this into your OpenClaw agent, Claude Code, or any coding agent with shell access:

Install https://github.com/oddessentials/odd-bot-research into my OpenClaw.
Follow https://oddessentials.github.io/odd-bot-research/AGENT-SETUP.md exactly.
Ask me for API keys before writing .env. Run ./install.sh --dry-run before ./install.sh.

Install in five commands

git clone https://github.com/oddessentials/odd-bot-research.git
cd odd-bot-research
cp .env.example .env     # fill in the keys you have; all optional
./install.sh --dry-run   # validates the config patch, changes nothing
./install.sh && openclaw gateway restart

Then ask your agent: research "What are the latest breakthroughs in protein folding?"

Requirements

An onboarded OpenClaw: ~/.openclaw/openclaw.json exists and openclaw --version works.
Node 20+, git, python3, and pipx on PATH. Missing pipx: brew install pipx && pipx ensurepath.
Optional API keys. With none you still get paper-search (arXiv, PubMed, Scholar), BCC Research, fetch, knowledge-graph-memory, arxiv-watcher, and Wikipedia.

What gets installed

Component	Searches	Key
`deep-research` skill	The orchestrator: plan, search, verify, synthesize, persist	none
exa	Semantic web search, deep researcher, company and people search	`EXA_API_KEY`
paper-search	arXiv, PubMed, Google Scholar, bioRxiv, medRxiv	optional
ieee-xplore	7M+ IEEE engineering and CS papers	`IEEE_API_KEY`
bcc-research	Market sizing, forecasts, industry analysis	none
firecrawl	JS-rendered scraping	`FIRECRAWL_API_KEY`
fetch, knowledge-graph-memory	URL to markdown; persistent findings	none
search-layer	Multi-engine search with dedup (Brave, Exa, Tavily, Grok)	per engine
last30days, x-research	Reddit, X, YouTube, HN, TikTok, Instagram, Bluesky discourse	`SCRAPECREATORS_API_KEY`, `XAI_API_KEY`
arxiv-watcher, wikipedia-oc, pubmed-edirect	arXiv, Wikipedia, PubMed via EDirect	none / `NCBI_API_KEY`
Built-in `web_search`	Brave Search, set as the provider	`BRAVE_API_KEY`

Sign-up links and free tiers: api-keys.md.

What the installer touches

~/.openclaw/mcp-servers/: the BCC proxy and a pinned build of ieee-mcp.
~/.openclaw/skills/deep-research/: the orchestrator.
~/.openclaw/workspace/skills/: pinned third-party skills.
~/.openclaw/openclaw.json: one validated patch adding env.* keys, mcp.servers.*, and Brave as search provider. A timestamped backup is written first.
~/.openclaw/workspace/TOOLS.md: a research block between odd-bot-research:start/end markers.

Every step is idempotent. Re-run ./install.sh after git pull to update. ./scripts/doctor.sh verifies without changing anything.

Verify

./scripts/doctor.sh
openclaw config get mcp.servers
openclaw skills list | grep -iE 'last30days|search-layer|x-research|arxiv|wikipedia'

Doctor flags the one known trap: a workspace skill that registers as deep-research (usually ClawHub's agent-deep-research) and shadows the orchestrator. Remove it and restart the gateway.

MIT licensed. Third-party components keep their own licenses and were each audited before inclusion; see audit-report.md.