Network of 7 AI tool nodes connected to a central amber anvil

Chapter 13

Multi-Agent Setup

One setup, all agents. Configure Plan Forge for 7 AI tools.

New here? What this chapter is about. Plan Forge isn't tied to one AI tool. Whatever you (or your team) already use, GitHub Copilot, Claude Code, Cursor, Codex, Gemini, Windsurf, the same plans, instructions, and reviewer agents work in all of them. This chapter shows how to install the right files for each tool. You don't need all of them; just pick the agent(s) your team uses.

What gets installed, native config files for each agent (e.g. CLAUDE.md, .cursorrules, AGENTS.md) so the agent reads Plan Forge's guardrails automatically.
Why it matters, you get the same architecture rules, same reviewers, same skills, no matter which AI you're talking to. Switching tools doesn't mean re-teaching the rules.
Default, GitHub Copilot files always install (Plan Forge's reference implementation). Add others with the -Agent flag.

Read this first if you haven't: the orchestration concepts this chapter assumes, intent lanes, quorum advisory, and the Forge-Master reasoning UI, are introduced in the Forge-Master (Deep Dive) sub-chapter and exposed on the Dashboard — Forge-Master tab. Skim either before wiring a second agent in if those terms are new.

One Setup, All Agents

Terminal

# Add all agent adapters at once
.\setup.ps1 -Preset dotnet -Agent all

# Or pick specific agents
.\setup.ps1 -Preset dotnet -Agent claude,cursor

Copilot files are always installed. The -Agent flag adds native files for other tools, each with all 18 guardrail files embedded, prompts as native skills/commands, and 21 reviewer agents as invocable procedures.

Feature Parity Matrix

Deepest integration with VS Code + GitHub Copilot — instruction loading, MCP tools, lifecycle hooks, Coding Agent dispatch, GHAS, Copilot Metrics, and dashboard sync. Other agents emulate the Copilot baseline; the amber rows below are Copilot-native surfaces with no equivalent on other tools. See Copilot Integration and Plan Forge on the GitHub Stack.

Feature	Copilot	Claude	Cursor	Codex	Gemini	Windsurf	Generic
Auto-loading instructions	✓ Native	✓ Emulated	✓ Emulated	⚠ Manual	✓ Emulated	✓ Emulated	✗
Pipeline agents	✓ 6	✓ Skills	✓ Commands	✓ Skills	✓ Commands	✓ Workflows	✗
Reviewer agents	✓ 21	✓ 21	✓ 21	✓ 21	✓ 21	✓ 21	✗
MCP tools	✓ 105+	✓	✓	⚠ Partial	⚠ Partial	⚠ Partial	✗
Skills / slash commands	✓ 14+	✓ Slash	✓ Commands	✓ Scripts	✓ Commands	✓ Workflows	✗
Full Auto execution	✓	✓	✓	✓	⚠	✓	✗
Lifecycle hooks	✓ 8 incl. LiveGuard	✓ Emulated	✗	✗	✗	✗	✗
Pipeline handoff buttons	✓ click	⚠ guided	⚠ guided	✗	✗	✗	✗
Memory bridge	✓ OpenBrain	✓ Native	⚠	⚠	⚠	⚠	✗
🐙 Copilot & GitHub-Native Surfaces
Auto-generated `copilot-instructions.md`	✓ `pforge sync-instructions`	—	—	—	—	—	—
Memory hints sync (`copilot-memory-hints.md`)	✓ `forge_sync_memories`	✓ MCP	✓ MCP	—	—	—	—
Copilot Coding Agent dispatch (cloud)	✓ `--worker copilot-coding-agent`	—	—	—	—	—	—
Cloud agent provisioning (`copilot-setup-steps.yml`)	✓ template	—	—	—	—	—	—
PR review delegation to `@copilot`	✓ `pforge github review delegate`	—	—	—	—	—	—
GHAS chain (CodeQL · secrets · deps · code review)	✓ `cloudAgentValidation`	—	—	—	—	—	—
SARIF → plan auto-conversion	✓ `pforge plan-from-sarif`	—	—	—	—	—	—
GitHub stack readiness check (8 primitives)	✓ `pforge github status`	—	—	—	—	—	—
Copilot Metrics API leaderboard	✓ `pforge github metrics`	—	—	—	—	—	—
Dashboard Settings → Copilot tab (preview + sync)	✓ native	—	—	—	—	—	—
VS Code MCP transport (`.vscode/mcp.json`)	✓ auto-generated	⚠ manual	⚠ manual	—	—	—	—
`AGENTS.md` contributor guide	✓ read	✓ native	✓ native	✓ native	✓ native	✓ native	✗

✓ = native / first-class · ⚠ = guided or partial · ✗ / — = not supported · MCP = with pforge-mcp/server.mjs running. The amber rows are Copilot-native surfaces with no equivalent on other agents today.

GitHub Copilot (Default)

Native integration. Instruction files auto-load via applyTo. Agents appear in the agent picker. Skills invoke via /slash-command. Hooks run automatically. This is the reference implementation, all other agents emulate this behavior.

Key file: .github/copilot-instructions.md

Claude Code

All guardrails embedded in a single CLAUDE.md file. Claude Code reads this automatically at project root. Includes 11+ skills as slash commands, full auto mode, and memory hooks.

Key file: CLAUDE.md

Setup

.\setup.ps1 -Preset dotnet -Agent claude

Cursor

Rules written to .cursorrules and .cursor/rules/*.mdc. Cascade integration loads rules automatically based on file patterns.

Key files: .cursorrules, .cursor/rules/

Codex CLI

Skills as executable scripts in .agents/skills/. Terminal-based execution with all pipeline steps available.

Key file: AGENTS.md

Gemini CLI

Guardrails embedded in GEMINI.md. Commands as .gemini/commands/*.toml files for /planforge-* invocations.

Key files: GEMINI.md, .gemini/commands/

Windsurf

Rules in .windsurfrules and .windsurf/rules/*.md with trigger frontmatter. Workflows mapped to Cascade integration.

Key files: .windsurfrules, .windsurf/rules/

Generic (Any AI Tool)

A single AI-ASSISTANT.md file with copy-paste guardrails. Works with ChatGPT, Ollama, or any tool that accepts text prompts.

Key file: AI-ASSISTANT.md

Cloud Agent

GitHub's Copilot cloud agent uses the same copilot flag, no separate adapter needed. Add copilot-setup-steps.yml to provision the agent's environment:

Terminal

cp templates/copilot-setup-steps.yml .github/copilot-setup-steps.yml

The cloud agent gets all guardrails, MCP tools, and pforge run-plan automatically.

OpenBrain: The Connective Tissue

Across all seven agents, one challenge remains: each tool starts each session with a blank slate. OpenBrain solves this by acting as a shared, persistent memory layer that every agent reads from and writes to, regardless of which tool authored the thought.

When Claude Code resolves an architectural ambiguity, that decision is captured as a thought. When you switch to Copilot the next morning, it retrieves that thought before writing a single line. When your team's Cursor instance encounters the same pattern, it inherits the same guardrails. The agents change; the institutional knowledge compounds.

How it works at the tool level: the Memory bridge row in the Feature Parity Matrix above shows each agent's integration tier. Copilot and Claude have full native integration; Cursor, Codex, Gemini, and Windsurf use the pforge recall CLI to inject context at session start. The Generic adapter includes copy-paste recall snippets.

For a deep dive into the three-tier memory architecture (in-RAM hub → local JSONL → pgvector semantic index), see Unified Memory Across Agents in Chapter 24.

See also: One Framework, Seven AI Agents, a practical walkthrough of how a mixed-agent team operates on a shared Plan Forge project without knowledge silos.

Spec Kit Interop

If you use Spec Kit for specifications, Plan Forge picks up where your specs end. The setup wizard auto-detects existing Spec Kit files and imports them as context. Extensions marked speckit_compatible work in both frameworks.

📄 Full reference: AGENT-SETUP.md on GitHub