Your AI says the work is done. LASDK makes it prove it.

LASDK, the Local Agentic Software Delivery Kit, turns any AI coding model into a governed software-development lifecycle that runs on your own machine.

LASDK does not upload your code, prompts or run evidence to Siklab — it runs entirely on your machine and has no telemetry you did not switch on. Note the honest boundary: LASDK drives your AI coding CLI, and that CLI sends project context to its own model provider under that provider’s terms. We never see it; your provider does.

LASDK is model-independent — it never talks to a model itself. It drives a supported AI coding CLI with your own login: today Claude Code and Codex. Whichever model those are pointed at is the model you get, and other harnesses can integrate through the LASDK run contract.

An active LLM subscription or API access (e.g. Claude or OpenAI/Codex) and a supported AI coding CLI installed (such as Claude Code or Codex). LASDK includes no model; it drives your CLI locally with your own login.

Diagnostics, docs, provider preflight, evidence verification, migration export and the judgement commands lasdk assess and lasdk certify are free — you can check any run’s evidence and gate your own repo without paying. Governed builds, critic runs, evidence generation and earned Done need an active subscription. A 24-hour Pro trial starts only after you activate it and a provider check passes, so setup time never burns it.

Partly, and the difference is the whole point. Graph engineering says to design the work around the model instead of pouring it into one long chat: split it into steps, pass state between them, and keep the writer separate from the checker — because a model grading its own answer will always like what it sees. LASDK agrees with every word of that. What it adds is enforcement. If the agent that wrote the code is also the one that reviewed it, the run doesn’t get to call itself done; the evidence is marked not independent and the gate refuses. Advice you can follow is not the same as a gate you can’t skip.

It also isn’t a workflow engine. LASDK doesn’t replace LangGraph, n8n or your agent harness — it sits above whichever one you use and decides whether the work they produced can be called finished.

In assist mode, it can: the agent chooses to call LASDK, so a rushed one can skip it; that’s a discipline you enforce. In governed mode, LASDK launches and controls the run itself: the agent runs with delivery credentials scrubbed and scoped to the workspace, can’t write to your tree (artifact-only), and can’t reach Done without passing the gate. It’s credential-custody + fail-closed delivery, not an OS sandbox. Use governed mode where it matters.