| name | xai-research |
| description | Research queries via xAI Grok — web search, X search, or plain chat. Triggers: "xai research" | "grok search" | "search xai" | "search x" | "search twitter" | "grok this" | "ask grok". |
| version | 0.1.0 |
| argument-hint | <query> [--web|--X|--all] |
| allowed-tools | Bash |
xAI Research
Query Grok (grok-4.3) via the llmCLI xAI forwarder. Supports plain chat, web search, and X (Twitter) search.
Let:
S := ${CLAUDE_PLUGIN_ROOT}/skills/xai-research/xai_research.py — research script
F := http://127.0.0.1:18645 — xAI forwarder (default endpoint)
C := ~/.roxabi/llmcli/xai-research.json — persistent config (optional)
Self-Check
Verify forwarder is reachable:
curl -s http://127.0.0.1:18645/health
logged_in: false or connection error → "xAI forwarder not reachable. Run systemctl --user start llmcli-xai-forwarder (M₁) or uv run python -m llmcli.proxy_forwarder._server & (dev). Then llmcli xai login if not authenticated." Halt.
Phase 1 — Parse Arguments
Parse $ARGUMENTS:
- Extract query string (everything that is not a flag)
- Detect mode flag:
--web | --X | --all | none
No query and no stdin → ask user "Enter your research query:".
| Flag | Tool sent to Grok |
|---|
| (none) | none — Grok answers from training data |
--web | web_search |
--X | x_search |
--all | web_search + x_search |
Phase 2 — Execute
Run S with --json for structured output:
uv run ${CLAUDE_PLUGIN_ROOT}/skills/xai-research/xai_research.py --json [--web|--X|--all] "<query>"
Parse JSON response. On error field → report HTTP code + detail + hint from table below. On success → extract and present output[].content[].text.
Phase 3 — Present
Present the response text directly. If sources were used (citations present in text), preserve them. If --raw was in $ARGUMENTS, rerun without --json and show full output.
Error Handling
| Error | Hint |
|---|
Connection error | Forwarder down — start with systemctl --user start llmcli-xai-forwarder |
HTTP 401 | Token expired — run llmcli xai login |
HTTP 422 | Wrong tool type or endpoint mismatch |
(no output_text) | Rerun with --raw to inspect raw response |
Safety
- ¬leak credentials — never echo api_key or access_token
- ¬modify config — suggest, don't write
- Research is read-only
$ARGUMENTS
