oracle 🧿 — Whispering your tokens to the silicon sage

Oracle gives your agents a simple, reliable way to bundle a prompt plus the right files and hand them to another AI. It currently speaks GPT-5.1 and GPT-5 Pro; Pro runs can take up to ten minutes and often return remarkably strong answers.

Two engines, one CLI

• API engine — Calls the OpenAI Responses API. Needs OPENAI_API_KEY.
• Browser engine — Automates ChatGPT in Chrome so you can use your Pro account directly. Toggle with --engine browser; no API key required.

If you omit --engine, Oracle prefers the API engine when OPENAI_API_KEY is present; otherwise it falls back to browser mode. Switch explicitly with -e, --engine {api|browser} when you want to override the auto choice. Everything else (prompt assembly, file handling, session logging) stays the same.

Note: Browser engine is considered experimental, requires an OpenAI Pro account and only works on macOS with Chrome. Windows/Linux browser support is in progress; until then, use --engine api or bundle files and paste manually. Your system password is needed to copy cookies. To skip Chrome/Keychain entirely, pass inline cookies via --browser-inline-cookies <json|base64> or --browser-inline-cookies-file <path> (fallback files at ~/.oracle/cookies.json or ~/.oracle/cookies.base64). API engine is stable and should be preferred.

Quick start

# One-off (no install)
OPENAI_API_KEY=sk-... npx -y @steipete/oracle -p "Summarize the risk register" --file docs/risk-register.md docs/risk-matrix.md

# Browser engine (no API key)
npx -y @steipete/oracle --engine browser -p "Summarize the risk register" --file docs/risk-register.md docs/risk-matrix.md

# Globs/exclusions
npx -y @steipete/oracle -p "Review the TS data layer" --file "src/**/*.ts" --file "!src/**/*.test.ts"

# Mixed glob + single file
npx -y @steipete/oracle -p "Audit data layer" --file "src/**/*.ts" --file README.md

# Dry-run (no API call) with summary estimate
oracle --dry-run summary -p "Check release notes" --file docs/release-notes.md

# Alternate base URL (LiteLLM, Azure, self-hosted gateways)
OPENAI_API_KEY=sk-... oracle --base-url https://litellm.example.com/v1 -p "Summarize the risk register"

# Inspect past sessions
oracle status --clear --hours 168   # prune a week of cached runs
oracle status                       # list runs; grab an ID
oracle session <id>                 # replay a run locally

How do I integrate this?

CLI (direct calls; great for CI or scripted tasks) - One-liner in CI: OPENAI_API_KEY=sk-... npx -y @steipete/oracle --prompt "Smoke-check latest PR" --file src/ docs/ --preview summary. - Package script: add "oracle": "oracle --prompt \"Review the diff\" --file ." to package.json, then run OPENAI_API_KEY=... pnpm oracle. - Don’t want to export the key? Inline works: OPENAI_API_KEY=sk-... oracle -p "Quick check" --file src/.

MCP (tools + resources; mix-and-match with the CLI sessions) - Run the bundled stdio server: pnpm mcp (or oracle-mcp) after pnpm build. Tools: consult, sessions; resources: oracle-session://{id}/{metadata|log|request}. Details in docs/mcp.md. - mcporter config (stdio): json { "name": "oracle", "type": "stdio", "command": "npx", "args": ["-y", "@steipete/oracle", "oracle-mcp"] } - You can call the MCP tools against sessions created by the CLI (shared ~/.oracle/sessions), and vice versa.

Highlights

• Bundle once, reuse anywhere — Prompt + files become a markdown package the model can cite.
• Flexible file selection — Glob patterns and ! excludes let you scoop up or skip files without scripting.
• Pro-friendly — GPT-5 Pro background runs stay alive for ~10 minutes with reconnection + token/cost tracking.
• Two paths, one UX — API or browser, same flags and session logs.
• Search on by default — The model can ground answers with fresh citations.
• File safety — Per-file token accounting and size guards; --files-report shows exactly what you’re sending.
• Readable previews — --preview / --render-markdown let you inspect the bundle before spending.

Configuration

Put per-user defaults in ~/.oracle/config.json (parsed as JSON5, so comments/trailing commas are fine). Example settings cover default engine/model, notifications, browser defaults, and prompt suffixes. See docs/configuration.md for a complete example and precedence.

Flags you’ll actually use

See docs/openai-endpoints.md for advanced Azure/LiteLLM configuration.

Sessions & background runs

Every non-preview run writes to ~/.oracle/sessions/<slug> with usage, cost hints, and logs. Use oracle status to list sessions, oracle session <id> to replay, and oracle status --clear --hours 168 to prune. Set ORACLE_HOME_DIR to relocate storage. Add --render (alias --render-markdown) when attaching to pretty-print the stored markdown if your terminal supports color; falls back to raw text otherwise.

Recommendation: Prefer the API engine when you have an API key (--engine api or just set OPENAI_API_KEY). The API delivers more reliable results and supports longer, uninterrupted runs than the browser engine in most cases.

Wait vs no-wait: gpt-5-pro API runs default to detaching (shows a reattach hint); add --wait to stay attached. gpt-5.1 and browser runs block by default. You can reattach anytime via oracle session <id>.

Testing

pnpm test
pnpm test:coverage

If you’re looking for an even more powerful context-management tool, check out https://repoprompt.com

Name inspired by: https://ampcode.com/news/oracle