Forge lets your AI question its own work. Independent agents research, draft, and critically review the complete copy against your original prompt. One objection sends it back for revision. You get the deliverable only when every reviewer approves it.
One prompt starts a council, not an autocomplete. Forge moves through discovery, grounded research, synthesis, and section-by-section drafting. Then every enabled agent audits the whole deliverable against the user's core ask. Any dissent triggers a complete revision and another independent vote.
critical-review.md; Codex keeps its delegated review trace in-session. No agreement means no fake “final.”
One command. Pick 2–5 providers with configured keys. Each becomes a participant named after the model it's running. The five debate roles (skeptic · pragmatist · analyst · advocate · contrarian) rotate between them at every phase — so the same question gets interrogated from five perspectives, each one coming from a different model every round.
forge debate "Should the municipality run the budget vote on-chain or on paper ballots?"Claude · Sonnet 4.5, Gemini · 2.5 Pro, GPT · 4o, DeepSeek V3, Ollama · Gemma 3 — each slot is a model, not a persona.
At each phase transition the RoleRotator shifts every role by one slot. Claude plays Skeptic in phase 1, Pragmatist in phase 2, Analyst in phase 3.
When an agent's role flips, a systemSuffix is injected with the new stance. The next response adopts the new perspective fully, no bleed-through.
CivicVote is a fictional product · decentralized voting for local municipalities. Below are four real decisions that team ran through Forge instead of chatting with one model for three hours each. For every one: the question, the mode, the artifact that landed on disk.
"Does the pain exist at a price someone will pay to fix?"
GO with three-line reasoning, plus a pivot shortlist if pilot #1 misses turnout. Skeptic forced the procurement-data question; Contrarian tried and failed to kill the idea.
"Nation-state? Incumbent? A coercive spouse? Name them before modeling them."
Three kill chains drafted, scored against real detection capacity. Ranked mitigation plan lands: kiosk mode, deniable re-vote, public tally feed, audit log with DIDs only.
"Cryptography is the part where silent bugs decide elections."
Findings report with file paths + severity · including a nonce re-use in vote/sign.ts:147 flagged HIGH. Architect, Security, Perf, and Tests all at the table.
"Pressure-test the pitch before the actual room burns it."
Investment memo ready for IC · verdict FOLLOW, with named next-diligence items. GP, Associate, LP proxy, and Founder all played out before the real meeting.
Same team. Different question. Same engine. CivicVote was the example. Your product is the point.
Phases, message limits, success criteria, loop detection · tuned for the decision type. Each row below is a real recorded session. Click to switch.
Press a in any running session. Swap each agent's model — Claude, Gemini, OpenAI, OpenRouter, Perplexity, Ollama — cycle providers, pause a loud voice, or force a specific perspective to speak next. Changes apply to the next turn without restarting the session.
Beyond the live panel, every flow has a CLI entry point:
forge auto "<request>" routes natural-language to the right mode;
forge pipeline "<spec>" chains modes with each phase's consensus
feeding the next; forge parallel splits into N sub-deliberations; forge
compress distills transcripts into handoff briefs; forge mcp runs
the blocking consensus tool for Claude Code, Codex, Cursor, and other MCP hosts.
Every live agent shows up · state, role, current model, pause badge. ← → cycle the model, p switches provider, space pauses, s forces a speak turn. Close with esc, changes apply immediately.
From Agent Control hit k · browse everything the loader found (skills/,
~/.claude/skills/forge/, plugin skills, skills.sh list output).
space toggles; the agent's next response incorporates the new knowledge
without a restart.
The deliberation engine enforces a specific kind of disagreement. Five culture-neutral reasoning stances, each with a job. No invented names, no backstories, no simulated humans · just structural roles that have to carry their own weight or the phase machine overrides them.
Run Forge directly in the terminal or Electron app, or install its host-native workflow in Claude Code or Codex. Claude Code runs the blocking Forge MCP engine; Codex delegates to three native host reviewers. Both stop on unresolved dissent and return the complete copy to the conversation only after unanimous approval.
# terminal + Electron git clone https://github.com/SaharBarak/forge.git cd forge && npm install && npm run cli:build && npm link forge # or: npm run electron:dev # Claude Code plugin claude plugin marketplace add https://github.com/SaharBarak/forge claude plugin install forge@forge # Codex plugin codex plugin marketplace add SaharBarak/forge codex plugin add forge@forge # then ask either host "Use Forge copywrite consensus for this landing-page brief…"