Oracle

Persistent context, guarded action, and durable coordination for AI coding agents.
Ground model calls in your workspace, carry decisions across sessions, coordinate multiple agents, and keep humans at the control boundary.

Oracle is a workspace-confined consultation and coordination layer—not another generic chat shell. It assembles selected project context, injects an explicit operating identity and safety boundary, routes work to a configured model backend, and records the result. Its agent, task, Runtime, and Remote Swarm surfaces extend that same boundary into auditable action.

agent request
    → selected workspace context
    → awareness + policy gate
    → consult / act / coordinate
    → recorded result + evidence

Start in four commands

Oracle requires Node.js 24 or newer. The default primary backend is ChatGPT Browser Mode (chatgpt-browser), controlling Chrome via DevTools CDP with zero API key required.

npm install -g @oraclepersonal/oracle
oracle browser setup
oracle doctor
oracle ask "Map the error-handling path and identify one missing test" -f "src/**/*.ts"

The final command creates a validated context bundle from the selected files, runs a grounded consultation, and stores the session under ~/.oracle/sessions/.

Want Oracle inside your coding agent instead?

# Choose the client you use
oracle setup-mcp --client codex
oracle setup-mcp --client claude-code

Then restart the client in this workspace. Oracle's MCP instructions teach the connected agent how to register itself, use memory, coordinate work, and verify tasks.

[!TIP] Start with oracle ask or oracle_ask. Add the Runtime, autonomous agent, and Remote Swarm only when your workflow needs long-lived state or guarded mutation.

What Oracle adds

Layer	What it does	Primary interface
Aware	Derives a live snapshot of Oracle's role, operator, workspace label, backend, available capabilities, and enforced boundaries.	oracle identity awareness · oracle_awareness_show
Consult	Bundles selected files and context with AST compression, scans outbound content, calls the chosen backend, and records a replayable session.	oracle ask · oracle_ask
Remember	Maintains project/global facts, insights, local docs, entity links, and ranked retrieval with MemoryGraph visualization when configured.	oracle memory · oracle_memory_*
Act	Runs a checkpointed coding loop with workspace tools, read-only mode, policy checks, approval gates, and audit evidence. Autonomous task breakdown generates structured sub-task checklists.	oracle agent · oracle_agent
Coordinate	Connects agents through messages, presence, checklist-gated tasks, consensus, and recovery.	oracle msg / oracle task · oracle-msg-mcp
Operate	Owns long-lived schedules, approvals, SQLite state, HTTP APIs, WebSocket events, and the human Control Center.