Ask Oracle

Oracle is an open-source CLI tool built by Peter Steinberger (steipete) that packages your prompt and repository files into a context bundle and sends it to frontier AI models for code review, architecture analysis, and debugging. It supports API mode with direct provider keys, browser automation mode via Chrome DevTools Protocol, and a manual copy-paste fallback. The project is actively developed under the MIT license with its source on GitHub.

What It Is

Oracle sits between your terminal and a roster of "mythical pro agents" — GPT-5.5 Pro (default), Gemini 3.1 Pro, Claude Opus 4.1, Claude Sonnet 4.6, and others — providing a single unified interface regardless of which provider you use. Rather than opening a chat window and pasting code manually, Oracle assembles a markdown bundle from file globs, applies size guards, and ships it to the model in one command. The result is a reproducible, session-tracked consultation that can be replayed, continued, or chained into multi-model advisory panels.

How the Engine System Works

Oracle offers three execution paths:

• API mode — uses OPENAI_API_KEY, GEMINI_API_KEY, or ANTHROPIC_API_KEY for direct provider calls; GPT-5 Pro runs detach by default and can be reattached via oracle session <id>.
• Browser mode — drives a Chromium-based browser over CDP, letting you use ChatGPT Pro or Gemini without spending API tokens; stable on macOS, functional on Linux and Windows with extra flags.
• Render/copy mode — assembles the bundle and copies it to the clipboard for manual paste when automation is blocked.

Multi-model runs (--models gpt-5.5-pro,gemini-3-pro,claude-4.6-sonnet) aggregate cost, token usage, and lineage across providers in a single command, with --allow-partial to recover successful outputs when one provider fails.

Session and Lineage Model

Every Oracle run is stored under ~/.oracle/sessions/<id>/ with machine-readable metadata, transcripts, and any generated artifacts. The oracle status command displays parent/child lineage trees so you can trace follow-up chains. The --followup flag continues a saved ChatGPT browser conversation or an OpenAI Responses API run by session ID or resp_… response ID. Browser sessions save transcript.md and artifacts including Deep Research reports and generated images.

Coding Agent and MCP Integration

Oracle is designed to be called from within other AI coding agents. It ships an oracle-mcp stdio server that plugs into any MCP-aware client — Claude Code, Cursor, Codex — via .mcp.json or mcporter configuration. A bundled Codex skill can be copied to ~/.codex/skills/oracle and referenced in AGENTS.md/CLAUDE.md. The chatgpt-pro-heavy MCP preset targets ChatGPT browser mode with GPT-5.5 Pro and Pro Extended thinking for a compact second-opinion workflow from within another agent session.

Update: v0.14.1

The latest release is v0.14.1, published on 2026-06-18, according to the GitHub repository. The project was created in November 2025 and has seen continuous releases tracked in a CHANGELOG. Recent additions include Gemini browser mode with image generation and editing, browser auto-reattach for long Pro runs, multi-turn browser follow-ups, Project Sources management, performance tracing (--perf-trace), and Azure OpenAI endpoint support. The repository lists 2,597 stars and 265 forks as of the last data snapshot.