Architecture
Architecture
Claude acts as the orchestrator. It reads your request, picks an expert, and delegates to a provider over MCP. Each provider reaches Claude Code differently:
- Codex (GPT) - no MCP server of its own. codex-cli ships none (the old
codex mcp-serversubcommand is gone;codex mcpmanages external servers), so the unifieddeliberationserver spawnscodex execdirectly fromcore/providers/codex.js. Advisory-only and single-shot: nothreadId, noworkspace-write. - Gemini - a bundled zero-dependency Node bridge (
server/gemini/index.js) wraps the Antigravity CLI (agy). - Grok (xAI) - a bundled zero-dependency Node bridge (
server/grok/index.js) talks to the xAI Responses API (/v1/responses) over HTTP. Advisory-only: it cannot edit files, but it can read attached files.
Responses are synthesized by Claude, never passed through verbatim.
End-to-end flow on a typical request:
text
You: "Is this authentication flow secure?"
|
v
Claude: detects a security question, selects the Security Analyst
|
v
+-------------------------------------+
| mcp__deliberation__ask-gpt / |
| mcp__deliberation-gemini__gemini / |
| mcp__deliberation-grok__grok |
| -> Security Analyst prompt |
| -> expert analyzes your code |
+-------------------------------------+
|
v
Claude: "I found 3 issues..." (synthesizes, applies judgment)- Each expert has a specialized system prompt (in
prompts/). - Claude reads your request, picks the expert, and delegates over MCP.
- Responses are synthesized, not passed through raw.
- Multi-turn conversations preserve context via
threadIdfor chained work, and implementation retries before escalating to you.