Skip to content

Agent factory · Setup

One setup, one workflow, one evidence-gated finish

ai-dlc pins every machine to one profile and gives agents one workflow in Claude Code or Codex. It closes a ticket only when the merged commit has a passing CI receipt.

Get started Suite overview

work finish refuses until the PR is merged, then verifies the CI receipt and closes issue 42. It opens on the finished session and then replays the typing.
work finish refused while the PR is open, then completed: the CI receipt passed and issue 42 closed. Watch a new project pass its required checks (20 s).

See it work

machine enroll previews and applies a profile; machine status reports the pinned commit and one unbound credential. It opens on the finished session and then replays the typing.
machine enroll applied the pinned profile; machine status shows one credential not yet bound

The work recording in the hero takes one reviewed record from work start to a closed ticket. GitHub and its CI are a local stand-in that speaks the provider contract. The gates check a real local merge and the receipt project check wrote.

What it looks like

The workflow plugin's skills, as Claude Code lists them, and the checks a project requires:

Claude Code lists the ai-dlc-workflow plugin: eleven skills and three hooks

ai-dlc project check --required in the demo project: a clean tree and every required check passing

Where it sits

Where ai-dlc sits: the developer side of the agent factoryDeveloper machine, localai-dlcmachine enrolla pinned profilework new → startrecord, issue, branchwork pr → project checka PR and a check receiptwork finishmerged, CI green, closedTeam harness repositoryharness.lockpins the workflowplugin (skills, guide,hooks): off in stable,on in candidateSource of truthGitHubissues, pull requests,a CI receipt per mergecommit, decisions indocs/decisions/Runhangarruns coding agentsin sandboxes withthe pinned harnessJudgeflightdeckcompares runs withand without theplugin; posts theverdict on the PRpublishpluginpinnedplugintrials,tracesissue,PR,receiptverdicton PRGovern: Tower clears risky agent actions with policy decisions, approvals and signed receipts

ai-dlc does not run agents in the cloud or judge harnesses. hangar renders and runs sandboxes, flightdeck evaluates, and ai-dlc authors the one optional plugin they pin. Read more in the hangar docs, the flightdeck docs and the example team harness, which pins this plugin (how to pin it).

Use cases

  • A new engineer is productive on day one. They enroll their laptop from the team's profile at a reviewed tag. They get everyone's runtimes, Claude Code and Codex settings and named credentials, with their own paths and accounts. machine status shows what drifted later.
  • An agent cannot declare victory early. A developer and their agent take an issue through work start and work pr; work finish refuses until the PR is merged into the target branch and that merge commit has a passing CI receipt, then closes the issue.
  • A team measures whether its workflow helps. The maintainer publishes the workflow plugin into the team harness, off in stable and on in candidate. flightdeck compares agent runs with and without it.

Quickstart

Needs git and curl on macOS or Linux (native Windows is not supported yet). These commands take a fresh machine to a project whose required checks pass.

git clone https://github.com/Sean-Koval/ai-dlc.git
cd ai-dlc
sh scripts/bootstrap.sh --source
export PATH="$HOME/.local/share/ai-dlc/bootstrap/bin:$PATH"
ai-dlc project init ~/demo-app
cd ~/demo-app
git init -q
ai-dlc project setup --root .
git add -A
git commit -qm "chore: initialize with ai-dlc"
ai-dlc project check --required

Get started covers existing work repositories and contributing; Configure a project covers GitHub and profiles.

Features

Feature What you get Read more
Machine enrollment Versioned profiles pin runtimes, tools, agent clients and named credentials; each machine keeps its own paths and accounts Machine enrollment
Project scaffolding project init / project adopt for generic, Python, Node or Rust repositories, with setup steps and required checks Get started
Workflow plugin Guidance, eleven skills and CLI-optional safety hooks for Claude Code, with a Codex form; off by default Workflow plugin
Work lifecycle Reviewed work records bind an issue, branch, specification and PR; work finish gates on the merge and its CI receipt Development workflow
Receipts project check --required runs the team's checks and writes a receipt bound to the commit Release verification
Documentation upkeep Catalog, impact review, dispositions and a docs gate; decision records with knowledge-note frontmatter Project documentation
Agent access ai-dlc mcp serve exposes reviewed work, documentation and knowledge tools to agents Workflow tool map
Providers GitHub Issues + Projects (default), Jira Cloud, OpenSpec, Obsidian Architecture
Question Page
How do I set up a new or existing repository? Get started
How does a piece of work go from ticket to finish? Development workflow
Which command or skill does each stage use? Workflow tool map
How is the system built, and what are its parts? Architecture
What is in scope, and what is frozen? ADR 0001, ADR 0002 and Roadmap
What was verified in the current release? Release verification