Products & tools

ForgeLoop

An engineering control plane that makes AI coding agents accountable for verified work.

Live product

A hosted delivery system that turns a labelled GitHub issue into planned agent tasks, runs them on infrastructure you control, and delivers a pull request only when your checks pass.

Visit ForgeLoop

View screenshots

ForgeLoop illustrated cover showing a task moving through code, verification, and a pull request with a repair loop
Illustrated project cover showing a task moving through code, verification, and a pull request with a repair loop; not an application screenshot.
THE PROJECT

A closer look at ForgeLoop.

ForgeLoop addresses the work around a coding agent: deciding what it may do, coordinating dependent tasks, checking the change, recovering from failures, and delivering the commit that was reviewed. Its web console and enrolled runners keep repository execution close to the infrastructure the developer controls.

The challenge

Writing code is one stage of software delivery. An automated change still needs a clear scope, a working branch, appropriate checks, and a deliberate review point. When several agents or machines participate, the system also needs to track ownership, dependencies, failures, and the identity of the commit that passed verification.

The product experience

ForgeLoop turns eligible GitHub issues into a task graph with dependencies, acceptance criteria, and file ownership boundaries. Runners claim tasks through leases and execute them in isolated Git worktrees. Changes move through integration, configured verification gates, bounded repair attempts, and independent review before delivery returns a pull request to the original repository.

  • An unattended run on a public repository took issue #10 to merged PR #11 with no human approval recorded, its required test gate passing, and about $0.22 in known model cost. The public demo replays that run from its recorded evidence.
  • Gate output, criterion coverage, and checksummed artifacts are tied to the exact commit under review, so a changed branch never inherits earlier evidence.
  • An October 2026 security audit fixed tenant, runner-sandbox, rate-limiting, and GraphQL-cost findings; runner 1.0.14 ships the hardening across Windows, macOS, and Linux previews.

Product & engineering

A GitHub App supplies selected repository access while the web control plane coordinates runs, membership, policy, and evidence. Runner-local tools execute commands, containers, browser checks, and configured MCP processes. Gate outputs and checksummed artifacts support review of the exact commit. Worktrees separate checkout changes; they are not a security sandbox. Human approval remains part of delivery policy, with guarded auto-merge available when explicitly enabled.

The decisions that shape the work

Separate coordination and execution

Give the console responsibility for delivery state and the runner responsibility for repository commands.

Verify a specific change

Tie acceptance coverage, gate output, and delivery checks to the commit under review.

Bound recovery

Feed failed checks into limited repair attempts and surface the need for attention when limits are exhausted.

Working within the constraints

  • Repository access and issue intake need explicit authorization before processing starts.
  • Execution and provider credentials remain on enrolled infrastructure controlled by the user.
  • A changed branch must not inherit verification evidence for an earlier commit.

Have a similar challenge?

For teams building agent-driven developer tools or operational automation, ForgeLoop illustrates the coordination, state, verification, and recovery systems needed around model execution.

Talk about your project
MORE WORK

More to explore.

All work
Mallace charcoal and crimson brand banner reading Find it. Prove it. Fix it.
Products & toolsLive product

Mallace

A security tool for Python, JavaScript/TypeScript, PHP and Ruby repositories on GitHub. It scans for serious bugs, proves each one with a failing regression test, and opens a pull request with the fix.

PythonSemgrepDockerGitHub ActionsGitHub AppOSVStripeNext.js
Puzzld banner reading Printable puzzles & worksheets on any topic, beside an example word search and maze
Products & toolsLive product

Puzzld

A puzzle maker for teachers, daycares, Sunday schools and families. Type a topic, pick an age, and Puzzld makes printable puzzles with answer keys: word searches, crosswords, mazes, dot-to-dots, reading passages and more. Teachers can assign them to a class and watch progress.

Next.jsTypeScriptFirebaseClaudeStripeVercel
Truss illustrated cover showing one agent runtime shared by terminal, editor, and desktop interfaces
Products & toolsOpen-source product

Truss

A provider-flexible coding agent with shared workspace context, permissions, and state across CLI, terminal, VS Code, and desktop clients.

TypeScriptAgent runtimeElectron
Hive website hero reading "One terminal. Many agents." with the Hive logo and a honeycomb backdrop
Products & toolsLive product

Hive

A Rust terminal multiplexer for AI coding agents. Claude Code, Codex and other agent CLIs run side by side in real terminals, managed by a local daemon that keeps them working when you detach.

RustTokioTerminal UILocal daemonGit worktrees
YOUR NEXT PROJECT

Let's build the software your work needs.

Start a conversation