← All projects

Developer tools / cross-platform

Forge

A development control plane for isolated agent tasks, cross-project context, human review, and desktop execution monitored from mobile.

BuiltAgent work runs in an isolated task worktree; approval, commit and push are separate human decisions, and workspace sync blocks a pull when both sides changed.

Verification notesImplemented · stabilization ongoing

System sketch
  1. 01Desktop + mobile web
  2. 02Context / isolated tasks / review
  3. 03Desktop runtime / explicit Git delivery
Simplified implemented boundaries. This is not a deployment topology or a product screenshot.
Forge interactive demoDemo data · no backend

Runs entirely in your browser. Opens with fictional sample data; Regenerate builds fresh rows from the project's schema. Clears on refresh.Open full screen ↗

Authorization in this portfolio

Execution service boundaries and GitHub authorization are separate from the portfolio's fixed command simulation. No terminal code is executed here.

01

Operational problem

Moving between a desktop, a phone, and a remote machine makes ordinary development work fragile. Projects, uncommitted files, and terminal sessions are easy to lose, and useful code is often faster to rewrite than to find again. Forge started as a way to keep project access, files, Git, and execution together instead of reconstructing that context by hand.

02

Technical constraints

A desktop host can touch the local filesystem. A mobile web client cannot. An execution session can outlive the page that started it. A GitHub login is a different credential from the Forge account. Sync has to show a conflict instead of silently picking a winner. Commands on a linked machine need an allowlist, timeouts, and path checks so a session does not become an open shell.

03

Architecture

The interface is a Next.js and React application. Local desktop work goes through Tauri commands backed by Rust. Web and mobile surfaces call scoped services. Workspace sync tracks revisions and blocks a pull when the same path changed in both places. Execution jobs live apart from the editor, with an allowlist, output limits, and persistent sessions. GitHub authorization stays distinct from the application session, including project-specific account selection on desktop.

04

Implementation

Sentinel aggregates project activity, runtime health, attention and review queues. Isolated agent task helpers create a session and worktree; the review rail separates Approve, Request Changes, Reject, Commit and Push. Durable-session storage records revisions, decisions and outcomes. Context descriptors distinguish editable task worktrees from read-only reference projects. This repository also contains project and file workflows, Git operations, device linking, and the execution service. Remote execution validates arguments, rejects shell metacharacters, and normalizes paths before use. Mobile editing is a web surface with its own session state. Model-assisted drafting sits behind provider and permission checks, and applying a change is a separate step from generating a draft. A shipped React Native or Expo client is not part of this repository.

05

Testing

Repository tests cover execution allowlists, path security, workspace revisions and conflicts, device tokens, and mobile GitHub session behavior. Those tests are evidence for specific boundaries. They are not a certification that every Forge workflow passes. Signed desktop release checks remain a separate gate.

06

Current status

Desktop and mobile-web implementations are present and still being stabilized. Sage can retrieve project context and draft inside those permission boundaries. It is not a fully trained autonomous coding agent. Mobile Sentinel is implemented as a web focus surface. Distributed orchestration includes foundation helpers and stubbed capability runners; broader cloud session wiring remains incomplete. A native mobile client is future work.

Project links

Public demo and source links have not been approved for this portfolio.

Continue exploring

All case studies →