Skip to content
kudamaPublic

About

AI Forge: an engineering platform for the Forge master agent and its scoped subagents, with a bounded local-model execution controller.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

AI Forge

AI Forge is an engineering workspace and execution platform for Forge, a master agent that coordinates specialist subagents on behalf of its operator. Forge owns planning, architecture, delegation, and final review; subagents carry out explicitly scoped tasks.

This repository currently implements a prototype controller for local language models, giving Forge a bounded way to delegate repository analysis and implementation proposals. Tool permissions and review remain explicit. The controller is one component of the agent system; it does not independently provide the master agent's reasoning or a complete autonomous subagent framework.

What it does

  • Coordinates local inference through Ollama, with configurable models per role.
  • Accepts authenticated tasks and keeps results isolated by caller.
  • Enforces explicit repository file manifests and read-only tool grants.
  • Provides deadlines, cancellation, execution limits, and persistent SQLite state.
  • Offers an optional local MCP bridge for desktop AI clients.
  • Supports reviewed reference examples and repeatable model evaluations.

Repository analysts produce source-grounded proposals. Implementers propose patches. The controller has no file-write, shell-command, or patch-application tool; an operator reviews changes and runs independent checks before applying them. A completed task is not a guarantee that its output is correct.

Getting started

The prototype uses Python 3.12 or newer and an existing Ollama installation. Start with installation and preflight to inspect the host and create a verified Python environment without replacing existing state. Then follow setup and operation for private credentials, source grants, installed models and controller startup. Credentials and task state stay outside version control.

For desktop integration, see the local MCP bridge. For startup and recovery on macOS, see service management. For verified restoration and authenticated LAN HTTPS, see private-network migration.

Architecture and validation

The independently installable shared agent core provides scoped read contracts for all six peers; see design and limits.

The planned UX Designer model evaluation compares local and hosted design helpers alongside engineering workers. Its requirements and handoff are captured; no evaluation result or deployed designer is claimed.

The work-order contracts and native enforcement define explicit assignment and result contracts, tested through the current controller compatibility path.

The controller and inference runtime are separate components and can run on separate hosts. The current validated environment is macOS; deployment to Ubuntu has not yet been validated. The default services bind to loopback. Remote access requires a separately reviewed authentication and transport setup.

Architecture · Validation · Reviewed examples and evaluation

Model outputs and repository content are untrusted. Runtime grants remain the source of authority; prompts cannot expand them. Do not expose raw model or controller listeners directly to the public internet.

The AI Forge backlog workflow uses GitHub Issues for scoped tasks, local-worker delegation, Forge review and verification.

License and support

AI Forge is available under the MIT License, including for commercial use. This prototype is provided as-is, with no guaranteed support, maintenance schedule, response time, or service availability. Third-party dependencies and model weights retain their own licenses.

About

AI Forge: an engineering platform for the Forge master agent and its scoped subagents, with a bounded local-model execution controller.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages