Many specialists. One coordinated system.

A local-first control plane for planning, executing, verifying, and governing artificial intelligence (AI) coding work.

Quick start · Capabilities · Visual tour · Near-term roadmap · Wiki

About

FORGE is a browser-based control room for AI-assisted software delivery. It connects projects, providers, coding agents, durable work packages, bounded execution, evidence, approvals, and GitHub workflows in one operator-controlled system.

The product goal is not “give a chatbot shell access.” FORGE separates model judgement from deterministic execution:

Human intent
  -> Architect plan
  -> durable work packages
  -> capability and Model Context Protocol (MCP) admission
  -> bounded context
  -> specialist handoff artifacts
  -> quality assurance (QA) / Reviewer / Security gates
  -> evidence, recovery, and GitHub handoff

FORGE is currently a single-operator beta. It can plan approved specialist packages and produce handoff artifacts for review. Specialist execution and file materialization are currently unavailable because Forge does not yet have an operating-system-enforced confined writer. The work-package and ACP flags are reserved settings; they cannot enable execution today. Direct host repository writes remain unavailable. It also stops short of autonomous commits, merges, broad live MCP authority, or unrestricted host control.

Why FORGE

Principle	What it means in practice
Human authority	Plans, risky capabilities, review gates, and final outcomes remain inspectable and controllable.
Deterministic boundaries	Models produce structured plans and requests; FORGE validates scope, policy, commands, files, and outcomes.
Bounded context	Specialists receive the minimum relevant repository evidence rather than unrestricted filesystem access.
Durable evidence	Tasks, attempts, work packages, runs, artifacts, approvals, and blocked reasons survive beyond a chat session.
Runtime-neutral agents	Direct model application programming interfaces (APIs), local models, the Codex command-line interface (CLI), and Claude Code can sit behind consistent FORGE contracts.
GitHub-native delivery	Structured issues, controlled handoffs, pull request (PR) contracts, and acceptance-criteria checks keep implementation traceable.

What FORGE Does Today

Plan and coordinate

• Create local or GitHub-backed projects.
• Configure cloud, local, and Agent Client Protocol (ACP)-backed providers.
• Ask the Architect to produce an implementation plan.
• Materialize plans into durable work packages, dependencies, capabilities, and review gates.
• Revise, approve, reject, retry, stop, and recover work from the dashboard.

Execute within policy

• Queue work through Redis and persist orchestration truth in PostgreSQL.
• Prepare specialist handoff artifacts for review after approval.
• Build bounded repository context packets for planning and handoff evidence.
• Broker MCP requirements through explicit admission and approval state.

Review and preserve evidence

• Store plans, execution output, command results, repository evidence, findings, and rework history as artifacts.
• Require QA and Reviewer gates, with Security gates for higher-risk work.
• Keep structured task attempts and blocked/recovery reasons.
• Validate GitHub issues before agent work.
• Generate controlled Claude Code or Codex handoff packages.
• Check pull requests against linked issue acceptance criteria.

How A Task Flows

Create task
  -> Redis wakes the worker
  -> Architect plans
  -> FORGE materializes packages and gates
  -> operator reviews the plan
  -> capability/MCP admission runs
  -> specialist handoff artifacts are prepared for operator review
  -> operator reviews and manually applies any separately supplied files
  -> QA / Reviewer / Security evidence is collected
  -> operator accepts, requests rework, or stops

The web app is the control plane. PostgreSQL stores durable state. Redis carries wake-up, retry, and dead-letter transport. The worker performs planning, handoff, execution, recovery, and evidence capture.

Quick Start

From the repository root:

bash scripts/install.sh

The installer prepares local services, creates ~/Documents/Forge/config/forge.env, installs web dependencies, prepares the database, and can configure a small local Ollama model.

Useful variants:

bash scripts/install.sh --check      # inspect readiness without changing the machine
forge upgrade                        # sync dependencies and migrations after pulling updates
FORGE_SKIP_OLLAMA=1 bash scripts/install.sh

Start FORGE:

forge

Open:

http://localhost:3000

The first account creates a password and, by default, a passkey. To use password only, set FORGE_PASSKEYS_ENABLED=0 in ~/Documents/Forge/config/forge.env before creating the first account.

For a no-cost plumbing test, run with the mock Architect:

cd web
FORGE_WORKER_MOCK_ARCHITECT=1 npm run dev

Try A Task

1. Open the dashboard.
2. Apply a provider preset or add a provider manually.
3. Create a project from a GitHub repository or local folder.
4. Create a focused task with testable acceptance criteria.
5. Wait for the Architect plan and work-package preview.
6. Review capability/MCP requirements and approve or revise the plan.
7. Inspect execution evidence, generated files, commands, and review gates.
8. Accept the result, request rework, or fix a blocked condition and retry.

Safety Model

FORGE is designed around explicit ceilings rather than presumed agent trust:

• project paths are validated before execution;
• context packets are bounded and inspectable;
• generated files are size/count limited and path checked;
• supported validation commands are allowlisted;
• MCP proposals, approvals, and effective grants are separate states;
• secrets are stored as encrypted settings or environment-variable names;
• prompts, logs, and evidence use redaction and bounded output;
• higher-risk work requires Security review;
• ACP adapters remain local processes and are not presented as operating system (OS) confinement;
• final authority remains with the operator.

The next deterministic-execution foundation is #201 — Operation Catalog and typed execution harness: agents select approved typed operations, while FORGE constructs, executes, and verifies the actual action.

Current Boundaries

Not built or intentionally deferred:

• broad live MCP tool/credential grants for specialists;
• general automatic branch, commit, PR, merge, or deployment authority;
• parallel specialist execution;
• fully autonomous QA, Reviewer, or Security gates;
• arbitrary model-authored shell execution;
• earned autonomy without verified historical evidence;
• the full dockable Forge Workspace shell and link graph.

Specialist execution is currently unavailable. These reserved settings do not override the missing confined writer; leave them unset or disabled:

FORGE_WORKFORCE_MATERIALIZATION=0
FORGE_WORK_PACKAGE_HANDOFF=0
FORGE_WORK_PACKAGE_EXECUTION=0
FORGE_ACP_WORK_PACKAGE_EXECUTION=0

Host repository writes are not an available execution path. Leave FORGE_HOST_REPOSITORY_WRITES unset or disabled. Setting it to an enable value does not make file materialization available; the request fails closed because path validation is not an operating-system sandbox.

Roadmap

The near-term order is deliberately reliability-first:

1. finish and prove MCP admission under Epic #172;
2. run focused end-to-end and failure testing;
3. close observed safety, recovery, and operator-blocking bugs;
4. add the deterministic Operation Catalog in #201;
5. normalize execution outcomes and stop reasons in #185;
6. continue the reliability and earned-autonomy work in Epic #184;
7. resume broad Forge Workspace expansion after the trust layer is proven.

See the near-term execution roadmap for exit criteria and the broader product roadmap for the long-form direction.

Visual Tour

Setup and provider configuration	Provider readiness
Architect plan and approval	Completed orchestration evidence

Core Vocabulary

Term	Plain-English meaning
Dashboard	The browser user interface (UI) where projects, providers, tasks, approvals, and evidence are managed.
Project	A local or GitHub-backed repository FORGE can reason about.
Task	A user objective submitted to FORGE.
Architect	The planning agent that turns intent into a bounded plan and work packages.
Work package	A durable, scoped unit of execution with role, capabilities, steps, and acceptance criteria.
Harness	A reusable execution overlay describing prompts, tools, inputs, outputs, validation, and provider preference.
Artifact	Persisted plan, evidence, output, finding, report, or log linked to a run.
Approval gate	A human or policy checkpoint before work can progress.
MCP admission	The policy decision that classifies and approves, defers, or blocks requested MCP capabilities.
ACP provider	A local coding-agent CLI connected through Agent Client Protocol.
Project Sentinel	Planned deterministic regression and workflow monitoring under Epic #184.
Operation Catalog	Planned typed execution surface where agents request approved operations rather than arbitrary commands.
Forge Workspace	Planned dockable workbench linking browser, repo, docs, terminals, GitHub, Notion, and task evidence.

Documentation

Guide	Purpose
Wiki overview	Plain-English product overview.
Operator guide	Install, configure, run, repair, deploy, and troubleshoot FORGE.
Developer guide	Web app, worker, database, tests, prompts, and implementation conventions.
Near-term roadmap	Current execution order and exit criteria.
Product roadmap	Broader beta, Workforce, and Workspace direction.
Visual identity	Logo meaning, components, motion, status, accessibility, and generated assets.
Design guide	Product model, UI principles, screenshots, and visual QA.
CLI architecture	forge command taxonomy and routing.
ACP and Zed connector	How FORGE connects to Codex CLI and Claude Code.
GitHub-native workflow	Controlled issue-to-handoff-to-PR workflow.
GitHub PR contract	Required PR structure and acceptance-criteria evidence.
GitHub agent run log	Durable GitHub workflow run state.
Forge Workspace roadmap	Future dockable workspace and context-linking plan.
Architecture decisions	Durable technical decisions and safety boundaries.

Repository Profile

Suggested GitHub About description:

Local-first control plane for planning, executing, verifying, and governing AI coding agents with bounded context and human approval.

Suggested topics:

ai-agents  coding-agents  agent-orchestration  local-first  nextjs
typescript  postgresql  redis  mcp  acp  developer-tools

The repository is public, but FORGE remains a rapidly evolving beta. Review the current boundaries and operator documentation before relying on it for important or sensitive work.