Kraft Lite

Install Kraft Lite and run a chain of work inside a single agent session, with no service, port, or database process. The chain has the same ordered nodes, gates, and fix-loop caps as Kraft. It is the attended half of Kraft: the chain runs in front of you and resumes across sessions, but it doesn't outlive your terminal. Lite is a Claude Code plugin, and the repository also ships a Cursor plugin manifest for it.

Unlike the kraft plugin, Lite needs neither the kraft binary nor an MCP server.

Before you start

• Python 3.10 or newer, and nothing else. Check with python3 --version.
• Claude Code.

Install

Lite ships in the same plugin marketplace as the kraft plugin:

claude plugin marketplace add itsOmidKarami/kraft
claude plugin install kraft-lite@kraft

Add --scope project to either command to keep it to one repo. Update with claude plugin update kraft-lite.

Run your first chain

In a session in the repo:

1. Run /kraft-lite:init, once per repo. It detects the test command and which installed skills can run each node, and writes an editable .kraft-lite/registry.yaml.
2. Run /kraft-lite:start and say what the work is. It makes a chain from the shipped one, unless you name another chain file, and runs the first node. It reports the chain's ID and where its state lives: bd, or a file under .kraft-lite/.
3. Run /kraft-lite:next to run the next node. It also resumes a node after a gate, a new session, or a cleared context.
4. At a gate, run /kraft-lite:gate. It presents what the gate needs decided, then records your approval or rejection.
5. Run /kraft-lite:status at any time to see which node is live, what blocks it, and how many fix attempts are spent.

Check it worked

After the install, run /kraft-lite:init in a repo, then /kraft-lite:status. Status reports that no chain exists yet. After you start one, it names the chain's first node (spec, in the shipped chain).

How Lite works

• Nodes. The shipped chain has ten nodes, from spec and plan through implementation and verification to opening a merge request, checking it, review by a person, merge and a post-merge watch, with four gates where a person decides.
• Who runs each node. Lite supplies the order, the gates, and the caps. The registry binds each node to a skill you already have, and Lite calls it. Lite doesn't supply a spec writer or a planner.
• Freezing. A chain freezes when it starts. You can't reorder, add, or remove its nodes. The registry is reread on every dispatch, so rebinding a node in it corrects a chain already in flight.
• State. Lite keeps state in beads (bd) when the repo has it, and otherwise in .kraft-lite/chain.jsonl. The format is the same bd export format either way, so adopting bd later is bd import.
• Several chains. One directory can hold several chains. With two or more unfinished chains, each command needs --chain-id <id>.

What Lite doesn't do

Lite has no unattended execution, no web board, no waiting on CI while you are away, no spend caps, and no cross-repo search. For those, use Kraft.

Next steps

• The Kraft Lite README: the source of truth for the chain's contents, what runs each node, and state handling. It changes faster than this page.
• Install: the service with a web board, spend caps, and chains that run unattended.