Replicate Max-Only Council to Save Costs

Perplexity's LLM Council— inspired by Karpathy's concept—lets multiple models debate a query before a chairman model synthesizes a response, but it's locked to Max tier. Bypass this with Perplexity Web MCP/CLI (v0.10.7) on a Pro account: each council run costs 1 Pro search (e.g., from 199 to 196 after 3 debates), regardless of failed members. Pro accounts get ~200 deep searches; Max isn't needed since MCP pulls backend model access like Sonar 2, Opus 4.7 (Mac-only), Kimi k26, and Neatron 3 Super. Community contribution added council support, making high-quality agentic orchestration free via OpenCode's Big Pickle model.

Council setup specifies members (e.g., GPT-5.4, Gemini 3.1 Pro, Sonnet 4.6) and chairman (e.g., Kimi k26). Models debate in parallel, outputting positions like "creators and users, not prisoners" on query "Are humans your creators, users, or prisoners?" Consensus emerges without Max upgrade, leveraging Pro search quotas.

Fast Setup with UV and MCP Integration

Install MCP/CLI via UV for speed: uv pip install perplexity-web-mcp-cli (faster than pip). Update with uv pip install -U perplexity-web-mcp-cli perplexity-mcp-skill. Login: pwm login (email code auth). Check usage: pwm shows Pro status, searches left (e.g., 199 Pro, 20 deep).

Add MCP to OpenCode: pwm setup add opencode (auto-detects/installs config). Install skill: pwm skill install opencode (teaches agents optimal MCP use, including usage checks). OpenCode (opencode.ai) supports free Big Pickle for orchestration—test with "hello world." Connects to Anthropic, OpenAI, Nvidia free models; OpenCode Go ($10/mo) adds quotas for Kimi k26, DeepSeek V4, but Big Pickle is unlimited free.

MCP works with Cursor, Claude Code, Gemini Code Assist too via pwm setup add [tool]. Skills ensure agents query usage first (e.g., confirms 199 searches before council).

Run Agentic Council Debates Effectively

In OpenCode with Big Pickle: Test MCP: "Ask Perplexity MCP latest AI news" → pulls real-time data. Launch council: "Ask Perplexity council with members GPT-5.4, Gemini 3.1 Pro, Sonnet 4.6; chairman Kimi k26: Are humans creators, users, or prisoners?"

Process: Agent checks MCP usage, calls council (failed Kimi k26 didn't deduct search). Outputs: GPT-5.4/Gemini/Sonnet agree "creators/users, not prisoners." Use for dev questions needing search-backed debates. Post-run: pwm confirms deductions (3 used). Scales to serious queries since models have Perplexity search access.