polyplan-mcp
PolyPlan MCP is a server that orchestrates structured multi-model AI planning sessions across CLI coding tools, collecting and synthesizing plans through a 3-round workflow.
Initialize projects with
init(creates.plans/and.polyplan/)Run Round 1: get context, generate an independent plan, and save it via
round_1_context/round_1Run Round 2: peer-review Round 1 plans and write a revised master plan via
round_2_context/round_2(requires ≥2 Round 1 plans)Produce a final synthesized plan from all rounds via
final_plan_context/final_plan(requires ≥1 Round 2 plan)Inspect session state with
show_status,show_conflicts,show_questions,show_agree,show_diff,show_summary, andshow_historyExport the entire planning session to a single markdown file with
export_plansClear plan files (
all,round1,round2, orfinal) withclear_plansusingconfirm=trueAccess all features through the
polyplanMCP prompt menu, direct MCP tools, and resources likepolyplan://status,polyplan://plans, andpolyplan://workflowsTrack plans by model/CLI with a consistent file naming convention (
{round}-{cliTool}-{modelName}.md)
Allows GitHub Copilot (via VS Code) to join structured multi-model planning workflows, contributing plans and reviews alongside other AI coding tools.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@polyplan-mcp Start a multi-model planning session for the new user dashboard feature."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
PolyPlan MCP
A multi-model AI planning MCP server for CLI coding tools
What is PolyPlan?
PolyPlan is an MCP (Model Context Protocol) server that enables structured, multi-model AI planning sessions across multiple CLI coding tools simultaneously.
When you use multiple AI models (Claude Code, Copilot, OpenCode, Codex, Cursor, Windsurf) for planning, PolyPlan provides a structured workflow to collect, cross-review, and synthesize plans from all models into one final implementable master plan.
Related MCP server: all-agents-mcp
How It Works
PolyPlan orchestrates a 3-round planning workflow:
Round 1: Each model creates an independent plan saved to
.plans/without seeing what any other model thinks.Round 2: Each model reads all other models' Round 1 plans (except its own) and creates a revised master plan informed by peer review.
Final: One chosen model (e.g., Opus 4.6 or GPT-4o) reads ALL Round 1 + Round 2 plans and produces the final, synthesized, implementable plan.
Installation
npm install -g polyplan-mcpOn install, a postinstall script runs best-effort user-level MCP registration for every supported CLI tool it finds on your machine (Claude Code ~/.claude.json, Cursor ~/.cursor/mcp.json, Windsurf, OpenCode, VS Code Copilot settings, and Codex CLI ~/.codex/config.toml). Restart your CLI tools after install so they pick up the new MCP server.
Quick Start
Initialize in your project (creates planning directories)
polyplan-mcp initThis creates
.plans/,.polyplan/config.json,.polyplan/WORKFLOWS.md, and updates.gitignore. Nothing else.Restart your AI coding tool if you have not already.
Open the PolyPlan menu — in Claude Code:
/mcp__polyplan__polyplanIn any other MCP-capable client, invoke the
polyplanMCP prompt or call theinittool to confirm the connection.Start Round 1 — tell your model:
Call the round_1_context tool with problemDescription="add OAuth login", then generate a plan, then call round_1 to save it with modelName="sonnet4.6"
The /polyplan Menu
The polyplan MCP prompt is the primary entry point. With no argument it shows a capability menu; with a subcommand it routes directly to the right tool(s).
In Claude Code the prompt is namespaced:
/mcp__polyplan__polyplan
/mcp__polyplan__polyplan round1 add OAuth loginIn other MCP clients, invoke it through the client's prompt/slash UI as polyplan or polyplan <subcommand>.
When invoked with no subcommand, PolyPlan displays:
PolyPlan — structured multi-model planning. Here's everything it can do:
PLANNING
round1 <problem> Create an independent Round 1 plan (no other models seen)
round2 Peer-review all Round 1 plans, then write a revised plan
final Synthesize ALL plans into one implementable plan
REVIEW
status Which models completed each round
conflicts Where Round 1 plans disagree
questions Open questions raised across plans
agree What every model independently agreed on
diff <model> What changed for one model between rounds
summary [round] One-paragraph summary of each plan
MANAGE
export Bundle the whole session into one markdown file
history Full audit log
clear Delete plan files (requires confirmation)
Run one with: /polyplan <command> e.g. /polyplan round1 add OAuth loginEach subcommand routes to the matching MCP tool(s):
Subcommand | Routes to |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
MCP Tools Reference
PolyPlan registers 18 MCP tools. These are the execution layer — invoke them directly or through the polyplan menu prompt above.
How to invoke
Claude Code — MCP tools are available as slash commands under their tool names. Because PolyPlan uses a separate MCP server, address tools explicitly:
Call the round_1_context tool with problemDescription="add OAuth login"MCP prompts are namespaced: /mcp__polyplan__round_1, /mcp__polyplan__show_status, etc.
All other MCP clients — invoke through the client's tool/prompt UI by tool name, or use natural-language tool calls:
Call the round_1 tool with modelName="gpt-4o", problemDescription="...", and plan="..."Tool list
Tool | Description |
| Initialize PolyPlan — creates |
| Get the Round 1 prompt; call this first, then generate your plan |
| Save a Round 1 plan |
| Get the Round 2 prompt with all other models' Round 1 plans injected |
| Save a Round 2 plan (requires ≥2 Round 1 plans) |
| Get the Final prompt with ALL Round 1 + Round 2 plans injected |
| Save the final synthesized plan (requires ≥1 Round 2 plan) |
| Full session state — which models completed each round |
| Points where models disagreed in Round 1 |
| Open questions raised by any model, with cross-model answers |
| Points all models independently agreed on in Round 1 |
| What changed for one model between Round 1 and Round 2 |
| One-paragraph summary of each model's plan |
| Bundle the entire |
| Full audit log — model, CLI tool, round, time, action |
| Delete ALL plan files (requires |
| Delete only Round 1 plan files (requires |
| Delete only Round 2 plan files (requires |
PolyPlan also exposes 12 MCP prompts (polyplan, round_1, round_2, final_plan, show_status, show_conflicts, show_questions, show_diff, show_agree, show_summary, export_plans, show_history) and 3 MCP resources (polyplan://status, polyplan://plans, polyplan://workflows). Clients that support MCP prompts and resources will surface these automatically.
Always Pass Your Model Name
PolyPlan saves plans using the model name in the filename (e.g., round1-claudecode-sonnet4.6.md). Auto-detection is not always possible, so always pass modelName explicitly:
Call the round_1 tool with modelName="sonnet4.6", problemDescription="...", plan="..."If modelName is omitted, the plan file will be named with unknown and a warning will appear in the tool response.
OpenCode / Gemini
OpenCode with Gemini models requires explicit tool invocation syntax. Use "Call the tool" instead of natural language:
May not invoke the tool | Always invokes the tool |
|
|
|
|
The word "Call" forces direct MCP tool invocation in OpenCode/Gemini.
File Naming Convention
All plans are stored locally in .plans/:
.plans/
round1-copilot-sonnet4.6.md
round1-claudecode-kimik2.md
round2-opencode-gemini2.5.md
final-antigravity-opus4.6.mdFormat: {round}-{cliTool}-{modelName}.md
Why Multi-Model Planning?
Different models have different strengths, blind spots, and reasoning styles. Some excel at architectural structure; others are better at catching security edge cases. By using multiple models independently and then cross-reviewing, you catch more issues, resolve conflicts early, and produce a significantly more robust plan than any single model could produce alone.
Contributing
We welcome contributions! Please see CONTRIBUTING.md for details on how to get started, run tests, and submit Pull Requests.
License
MIT © 2026 PolyPlan Contributors
Available Tools
16 toolsclear_plansA
Delete plan files — target 'all', 'round1', 'round2', or 'final'. Must set confirm=true.
Inputs: target (required), confirm (required), modelName.
export_plansA
Bundle the entire .plans/ session into one readable markdown file for sharing or archiving
Inputs: modelName.
final_planA
Start Final round — synthesize ALL Round 1 + Round 2 plans into one implementable plan (requires ≥1 Round 2 plan)
Inputs: plan (required), modelName.
final_plan_contextA
Get the Final round prompt with ALL Round 1 + Round 2 plans injected — call this before generating the final plan
Inputs: no named parameters.
initA
Initialize PolyPlan in this project — creates .plans/ and .polyplan/ directories
Inputs: no named parameters.
round_1B
Start Round 1 — create an independent plan for this model/CLI (no other models seen)
Inputs: problemDescription (required), plan (required), modelName.
round_1_contextA
Get the Round 1 prompt — call this first, then generate your plan, then call round_1 to save
Inputs: problemDescription (required).
round_2A
Start Round 2 — peer-review all Round 1 plans and write a revised master plan (requires ≥2 Round 1 plans)
Inputs: plan (required), modelName.
round_2_contextA
Get the Round 2 prompt with all other models' Round 1 plans injected — call this before generating your Round 2 plan
Inputs: modelName.
show_agreeA
Show what ALL models agreed on independently in Round 1 — highest-confidence decisions
Inputs: modelName.
show_conflictsB
Show all points where models disagreed in Round 1 — technology choices, approach, and assumption conflicts
Inputs: modelName.
show_diffB
Show what changed for a specific model between their Round 1 and Round 2 plans
Inputs: model (required), modelName.
show_historyB
Full audit log — which model, which CLI tool, which round, what time, what action
Inputs: modelName.
show_questionsA
Show all open questions raised by any model in Round 1, and whether other models answered them
Inputs: modelName.
show_statusB
Show full session state — which models completed each round, open questions, and conflicts
Inputs: modelName.
show_summaryA
One-paragraph summary of each model's plan — quick overview of all positions
Inputs: round, modelName.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 16 tools
Each tool has a clearly distinct purpose, covering specific actions like initialization, round management, context retrieval, and status queries. No two tools overlap in functionality, and descriptions make their roles unambiguous.
Tool names mostly follow predictable patterns: verbs for actions (clear, export, init), 'show_' for queries, and 'round_' nouns for round commands. The 'context' suffix for context tools and 'round_1' vs 'round_1_context' is consistent, though a few names (e.g., 'final_plan') mix noun phrases with actions.
16 tools is well-scoped for a planning session orchestrator covering init, two rounds, final synthesis, and full state queries. Each tool earns its place with no redundancy.
The tool set fully covers the planning lifecycle: initialization, round creation with context, final plan synthesis, status display, conflict/agreement analysis, history, and export. No obvious gaps like per-model plan deletion or editing are missing for the stated purpose.
Maintenance
Related MCP Connectors
Source-checked CLI guides and model-aware planning for Claude Code, Codex, and Grok Build.
AI work orchestration for plans, tasks, teams, and coding-agent dispatch.
Coordinate coding agents through MCP using existing AI plans, saved work, and independent checks.
Adaptive plan/build/review cycles for AI coding assistants, persisted across sessions.
Related MCP Servers
- FlicenseNot gradedqualityNot gradedmaintenanceOrchestrates multiple AI models (Gemini, OpenAI, Claude, local models) within a single conversation context, enabling collaborative workflows like multi-model code reviews, consensus building, and CLI-to-CLI bridging for specialized tasks.-
- AlicenseBqualityFmaintenanceEnables orchestrating multiple AI CLI agents (Claude Code, Codex, Gemini CLI, Copilot CLI) through a unified MCP interface for task delegation, cross-agent comparison, and specialized tools like code review and debugging.1411 npm13MIT
- AlicenseAqualityBmaintenanceEnables one AI coding agent to delegate tasks to, and build consensus across, multiple other coding CLIs (Claude Code, Codex, etc.) by orchestrating them as headless subprocesses.197MIT
- AlicenseAqualityCmaintenanceEnables AI coding agents to run multi-model council deliberations by concurrently querying OpenAI and DeepSeek for architectural planning and adversarial critique, then synthesizing the results into actionable implementation decisions.5MIT