Forge Engine

What the Forge Engine MCP server does

The Forge Engine MCP server gives AI coding agents access to the living design for a Forge Engine project. That design describes systems, their purpose, boundaries, acceptance criteria, milestones, and links to the files that implement them. Agents can retrieve task-specific context before working instead of relying on manually repeated instructions.

The server also coordinates work between agents and people. An agent can claim a task, making its ownership visible to others before implementation begins. It can report completed work, while design changes remain proposals until a human reviews and adopts them. Deletions follow the same review path.

Forge records decisions, rejected proposals, implementation status, and evidence related to acceptance criteria. It distinguishes unverified claims from work backed by named tests and from guarantees based on tests that have run successfully. Verification evidence can expire, and the system flags drift between the design and the files it describes.

How it works

The connector runs locally as the forge command and communicates with an MCP client over the standard local process configuration. The hosted Forge application supplies the project data and authenticated operations; this repository contains the open-source connector and runtime CLI rather than the hosted application itself.

Authentication starts with an Account Key generated in Forge under Connect Agent. Running forge login exchanges that key for a device-specific token. The Account Key is not stored on disk, and devices can be revoked from Forge. After login, the MCP client starts forge mcp for each project folder without placing a key in its configuration.

A Forge project can also expose a public, read-only design link. That path does not use the MCP connector and cannot submit proposals or expose member activity. The published document is compiled on each read and includes the project overview, systems, acceptance information, and milestone plan.

Setup and configuration

Forge Engine MCP server setup requires Node.js 18 or later and the @forgeengine/cli package. Install it globally with:

npm i -g @forgeengine/cli
forge login

Then add this MCP server entry to the client configuration:

{ "command": "forge", "args": ["mcp"] }

Restart the client after saving the configuration. A global installation is not required. Using npx instead downloads and caches the package when needed:

npx -y @forgeengine/cli login

The corresponding client command is:

{ "command": "npx", "args": ["-y", "@forgeengine/cli", "mcp"] }

The npx option can add startup time and needs network access on its first use. Both installation methods share the stored device session.

Tools and capabilities

The Forge Engine MCP server supports workflows for:

• Reading the relevant project and system context for a task.
• Viewing system boundaries and associated implementation files.
• Claiming tasks so concurrent agents can see ownership conflicts.
• Submitting proposals for additions, edits, or deletions without changing the live plan directly.
• Reporting implementation results and recording decisions or rejections.
• Checking acceptance evidence, test-backed verification, and design-code drift.

These capabilities suit teams that run several coding agents in parallel or need non-coders to review the same plan used by agents.

Limitations and notes

The MCP loop requires installation and authentication. Public design links provide reading only; they do not replace authenticated MCP access for proposals or other changes. ChatGPT is documented as unable to open these public links directly, although the design can be copied for manual pasting. The connector repository is MIT licensed, while the hosted Forge application and hosted MCP server are maintained separately.