Oracle — OpenClaw Skill

Delegate complex questions to other AI models by bundling prompts with file context.

Development Vetted

What This Skill Does

Oracle is a meta-AI skill that lets your OpenClaw agent get a second opinion from another AI model. It bundles your prompt together with selected files from your codebase into a single "one-shot" request that can be sent to GPT-5.2 Pro, Gemini, Claude, or other models. The output is treated as advisory -- your agent verifies the answer against your actual code and tests before acting on it.

The primary workflow uses browser automation with GPT-5.2 Pro in ChatGPT. This is the "long think" path where complex problems get 10 minutes to an hour of deep reasoning. Oracle manages sessions automatically, so if a query takes too long, you can detach and reattach later. It also supports API mode for faster responses from any OpenAI-compatible endpoint, and a manual paste fallback that copies the prompt to your clipboard.

Oracle is designed for hard problems: architectural decisions, debugging complex race conditions, security reviews, or getting tradeoff analysis from a different model's perspective. It supports glob-based file selection, exclusion patterns, dry-run previews with token estimates, and readable session slugs. Never attach secrets by default -- the skill filters .env files and auth tokens. For everyday coding tasks, the Coding Agent skill is more appropriate.

Example Prompts

Ask GPT-5.2 Pro to review our authentication flow -- include src/auth/** and explain the threat model

Get a second opinion on this architecture decision -- bundle the README and src/index.ts and ask for 3 alternatives with tradeoffs

Do a dry run first to see how many tokens this query will use before sending it

Send this debugging question to GPT-5.2 Pro with all the files in services/proxy/ and the error log I just pasted

Ask Oracle about the best way to handle database migrations with our current schema -- attach the migration files

Check the status of my running Oracle sessions and reattach to the security review one

Requirements

Binary dependency: oracle

  • Install: npm i -g @steipete/oracle
  • Browser mode: Requires a logged-in ChatGPT session (default)
  • API mode: Requires OPENAI_API_KEY or compatible endpoint

Setup on KiwiClaw

Oracle is pre-installed on all KiwiClaw tenant machines. For browser mode, configure your ChatGPT session in the KiwiClaw dashboard. For API mode, set your OPENAI_API_KEY under environment variables. Oracle automatically picks API mode when a key is present, or browser mode otherwise.

Setup Self-Hosted

  1. Install Oracle: npm i -g @steipete/oracle
  2. Verify: oracle --help
  3. For browser mode: log into ChatGPT in your browser
  4. For API mode: export OPENAI_API_KEY="sk-your-key"
  5. Test with a dry run: oracle --dry-run summary -p "test question" --file README.md

Related Skills

  • Coding Agent -- for everyday coding tasks and PR reviews
  • GitHub -- manage repos where Oracle-reviewed code lives
  • Model Usage -- track costs from Oracle API queries
  • Session Logs -- review past conversations for context

FAQ

What can the Oracle skill do in OpenClaw?

Oracle bundles your prompt with selected files from your codebase and sends them as a one-shot query to another AI model like GPT-5.2 Pro. It supports browser automation and API modes, session management, dry-run previews, and file attachment with glob patterns.

Which AI models does Oracle support?

Oracle supports GPT-5.2 Pro via browser automation (default), and any OpenAI-compatible model via API mode. Browser mode also supports Gemini models. Use API mode for Claude, Grok, Codex, or multi-model queries. See our platform comparison for model availability.

How long do Oracle queries take?

Oracle queries via browser mode with GPT-5.2 Pro typically take 10 minutes to 1 hour for complex problems. The CLI manages sessions so you can detach and reattach later. API mode is faster but may have different capabilities.

Is the Oracle skill safe to use?

The Oracle skill has been security-vetted by KiwiClaw. It does not attach .env files or secrets by default. Always review what files are being sent with --dry-run before submitting. Sessions are stored locally.

Get a second opinion from any AI model

Bundle prompts with code context and delegate to GPT-5.2 Pro, Gemini, and more.