Saltar al contenido
TakoAPI
Volver a skills

agent-deep-research

por 24601

Search & Research
2.0kDescargas según clawskills.sh
3Estrellas
0Vistas
0 me gustaInicia sesión para dar me gusta a esta skill
—Aún sin valoraciones
Tu valoración

Resumen

Autonomous deep research powered by Google Gemini.

Descripción

Qué hace esta skill

Runs deep research queries through Google Gemini's research API without requiring the Gemini CLI. Supports grounding queries in local files via RAG by uploading them to ephemeral file search stores. Works as a universal skill across Claude Code, Amp, Codex, and Gemini CLI.

Calls the Gemini API directly via uv, so it works in any environment without installing the Gemini CLI or managing Python virtual environments.

Cuándo usarla

  • Researching a library or framework before adopting it in a project
  • Answering architecture questions grounded in local source files
  • Generating a security audit report from uploaded project documentation
  • Previewing API costs before running an expensive research task
  • Running async research queries from CI pipelines or agent workflows

Flujo de trabajo de ejemplo

ENTRADA

User asks: Analyze the authentication architecture of my codebase

AGENTE
  1. 1

    Run research.py start with --context ./src and --dry-run to preview upload costs and file list

  2. 2

    Start grounded research, uploading source files to an ephemeral file search store

  3. 3

    Poll for completion using adaptive intervals tuned to past research history

  4. 4

    Save the full report, metadata, and source citations to a structured output directory

  5. 5

    Return a compact JSON summary to the agent with the report path and duration

SALIDA

A directory containing report.md, metadata.json, sources.json, and interaction.json, with a compact JSON summary printed to stdout

Requisitos

  • One of GOOGLE_API_KEY, GEMINI_API_KEY, or GEMINI_DEEP_RESEARCH_API_KEY environment variable

Instalación

clawhub install 24601/agent-deep-research

Ejecuta el comando de la CLI o pega el prompt en Claude Code, Cursor, Codex o el agente de programación que uses.

Enlaces