#Try it
After installing (brew install steipete/tap/oracle or npm i -g @steipete/oracle), every consult is a one-liner.
# Browser path — no API key, drives ChatGPT directly (default: GPT-5.5 Pro).
oracle --engine browser --browser-manual-login -p "Review the storage layer for schema drift" --file "src/**/*.ts"
# API path — multi-model cross-check in one run.
oracle -p "Cross-check the data layer assumptions" \
--models gpt-5.5-pro,gemini-3-pro,claude-4.6-sonnet \
--file "src/**/*.ts"
# Manual fallback — assemble the bundle and copy it to your clipboard.
oracle --render --copy -p "Architecture review" --file "src/**/*.ts"
# Sessions you can replay or continue.
oracle status --hours 72
oracle session <id> --render
oracle --followup <id> -p "Re-evaluate with this new context" --file "src/**/*.ts"--render emits the assembled markdown, sessions persist machine-readable metadata, and progress stays out of saved answers so pipes remain usable.
#What Oracle does
- One CLI to a stable of pro agents. GPT-5.5 Pro (default), GPT-5.5, GPT-5.4 Pro, GPT-5.4, GPT-5.2 Pro, GPT-5.1 Pro, GPT-5.1 Codex, Gemini 3.1 Pro, Gemini 3.5 Flash, Gemini 3.1 Flash-Lite, Claude Sonnet 4.6, Claude Opus 4.1 — plus any OpenRouter id.
- Engines, plural. API mode for reliability, browser mode (Chrome over CDP) when you don't want to pay or want the Pro tier,
--render --copywhen neither is an option. - Multi-model in one run. Aggregate cost, token usage, and lineage across providers in a single command.
- Recoverable panels.
doctor --providers,--preflight,--route, and--allow-partialmake provider/key failures clear without losing successful model output. - Followups + lineage. Continue from any stored session id or
resp_…response id;oracle statusshows parent/child trees. - Sessions you can replay. Every run is stored under
~/.oracle/sessions/<id>/. Reattach to long browser runs without re-spending tokens. - Built for coding agents. Use it from Claude Code, Codex, Cursor, or any MCP host via
oracle-mcp. Plain stdout JSON envelopes for scripting. - Bundles, not chats. Globs + excludes + size guards +
--files-reportso you know exactly what is shipped to the model. - Traceable startup.
--perf-tracerecords startup and first-output timing when agent handoffs need performance proof.
#Pick your path
- Trying it. Install → Quickstart. Five minutes from
brew installto your first answer. - Choosing a model. The Mythical Pro Agents lineup covers when to reach for GPT-5.5 Pro vs. Gemini 3.1 Pro vs. Claude Opus, and what each costs.
- Wiring up an agent. Agents covers Claude Code, Codex, Cursor, and the
oracleskill. MCP plugs Oracle into any MCP-aware client. - Driving ChatGPT without keys. Browser mode walks through manual-login profiles, attach-running, remote browsers, and Deep Research.
- Long Pro runs. Sessions and the followup flow handle background runs, reattach, and lineage.
#Why "mythical pro agents"?
The frontier models marked "Pro" — GPT-5.5 Pro, Gemini 3.1 Pro, Claude Opus, OpenAI Deep Research — are slow, expensive, and gated behind ChatGPT Plus / Pro tiers, separate APIs, or per-token bills that scale fast. Oracle is the single entry point: one config, one session store, one set of flags. Bundle the right files, ask the right model, get a second opinion without remembering which provider charges for what.
#Project
Active development under MIT. The changelog tracks recent releases. Source on GitHub. Not affiliated with OpenAI, Google, or Anthropic.