Skip to content

agents.md

Latest commit

 

History

History
116 lines (76 loc) · 5.16 KB

agents.md

File metadata and controls

116 lines (76 loc) · 5.16 KB
Raw
title Coding Agents
description Use Oracle from Claude Code, Codex, Cursor, and any other coding agent — as a CLI, as an MCP server, or as a one-shot skill.

Oracle is built to be called by coding agents as much as by humans. The flow is always the same: the agent gathers context, hands the bundle to a stronger Pro model, gets a second opinion back.

The 30-second wiring

Drop this into the project's AGENTS.md or CLAUDE.md:

- Oracle bundles a prompt plus the right files so a Pro model (GPT-5.5 Pro,
  Gemini 3 Pro, Claude Opus) can answer with real repo context. Use when stuck,
  debugging hard bugs, doing architecture review, or cross-validating a plan.
- Run `npx -y @steipete/oracle --help` once per session before first use.

That's enough for most agents to discover and use Oracle correctly. The patterns below cover the deeper integrations.

Claude Code

As an MCP server (recommended)

oracle bridge claude-config --local-browser > .mcp.json

That writes a .mcp.json configured for the local browser path, so Claude Code can call oracle.consult and oracle.sessions without any API keys. Use the MCP consult tool with preset: "chatgpt-pro-heavy" for ChatGPT GPT-5.5 Pro with Pro Extended thinking. Add dryRun: true to inspect the resolved bundle before sending.

See MCP for connection details and other clients.

As a skill

Copy the bundled skill into ~/.claude/skills/:

mkdir -p ~/.claude/skills
cp -R skills/oracle ~/.claude/skills/oracle

Then reference oracle in CLAUDE.md. Claude Code will load SKILL.md whenever the trigger conditions match (debugging, refactor, design check).

As a slash command

Many users alias Oracle behind a custom /consult slash command that wraps npx -y @steipete/oracle --engine browser …. Pair with --browser-tab current to keep all consults in one ChatGPT conversation.

Codex

Copy the same skill into the Codex skills folder:

mkdir -p ~/.codex/skills
cp -R skills/oracle ~/.codex/skills/oracle

Then reference it in AGENTS.md. Codex will pick it up automatically.

For Codex slash prompts, drop a wrapper in ~/.codex/prompts/oracle.md that calls Oracle with your preferred defaults (engine, model, follow-up flags).

Cursor

Cursor speaks MCP. Drop a .cursor/mcp.json like:

{
  "oracle": {
    "command": "oracle-mcp",
    "args": []
  }
}

Or use the one-click install. The oracle source then shows up in Cursor's MCP picker.

Generic CLI usage from any agent

When the agent has shell access, the simplest hand-off is the bundle-on-clipboard fallback:

oracle --render --copy -p "$TASK" --file "$RELEVANT_FILES"

…then the agent (or a human) pastes into whichever Pro model they have access to. No keys, no MCP, works everywhere.

For autonomous dry-runs, use the JSON preview to inspect the resolved bundle before spending model time:

oracle --dry-run json --model gpt-5.5-pro -p "$TASK" --file "$RELEVANT_FILES"

Completed runs persist answers, usage, cost, session ids, model choices, and lineage under ~/.oracle/sessions/<id>/. Exit code is non-zero on failure.

Multi-agent shared profile (browser mode)

When multiple agents share one signed-in Chrome profile (the manual-login workflow), Oracle coordinates browser tab slots so parallel runs queue instead of crashing. Tune with:

  • --browser-max-concurrent-tabs — default 3 simultaneous tabs.
  • --browser-profile-lock-timeout — wait for the profile lock before sending.
  • --browser-reuse-wait — wait for a shared Chrome profile before launching.

For the most reliable shared setup: run one signed-in Chrome with remote debugging, point all callers at it via --remote-chrome <host:port>. See Browser Mode.

Cost / safety hygiene

  • Always preview Pro runs. --dry-run summary --files-report before a Pro API call on a large bundle. Token counts are a close-enough proxy for dollars.
  • Cap file size. ~/.oracle/config.json → maxFileSizeBytes, or ORACLE_MAX_FILE_SIZE_BYTES. Default is 1 MB per file.
  • Excludes are your friend. --file "src/**" --file "!**/*.test.ts" --file "!**/*.snap" cuts most fixtures.
  • API mode runs cost real money. If your agent runs Oracle autonomously, scope it: pin --model, set --timeout, and review the session log. Many users gate API mode behind explicit user consent and let browser mode run free.

Patterns that work

  • Stuck → Oracle. When the agent has been spinning on the same bug for 3+ turns, hand the failing test plus the involved files to GPT-5.5 Pro. It often spots the issue in one round.
  • Plan → Oracle → execute. Draft the plan, ask Claude Opus or Gemini 3 Pro to challenge it, then implement.
  • Refactor → cross-check. After a non-trivial refactor, send the diff plus the spec to a different provider than the one that wrote the diff. Catches drift fast.
  • Followup chain. Use --followup <id> to keep one Pro session alive across iterations rather than re-bundling the whole repo every time. See Followup.