Local-first · a Kaltstart project

Vibe-code fast. Ship code that's actually done.

AI agents love to say “done” when they're not. Agentklar tracks the work, runs your real checks, keeps the proof — and lets only you sign off.

Get started See all features →

Try it — run the task, then finish what the agent can't.

agentklar task AK-142

What you get

Trust what your agent hands you.

Proof, not vibes

Agentklar runs your project's own checks and keeps the evidence — command, exit code, logs. No more “looks done to me.”

Only you call it done

The agent can claim, submit, and pass QA. It cannot approve its own work — that's your call, from your own account.

Bring your own agent

Codex, OpenCode, Gemini CLI — your model, your machine. Agentklar never holds your keys or picks your model.

Never lose the thread

Work lives in a local database, not the chat. Any session — today or next week — resumes exactly where the last one stopped.

See all features →

Get started

Build it, run one task end to end.

One Go binary. No installer, no Docker, no sudo. Here is the whole path.

Go 1.25+gitmacOS or Linux

1Build

git clone https://github.com/kaltstart-co/agentklar
cd agentklar && go build -o agentklar ./cmd/agentklar

2Create a taskyou set the goal

./agentklar init
./agentklar task new AK-1 Fix the parser \
  --criteria "handles empty input;tests pass" --verify "go test ./..."
./agentklar task ready AK-1

3Let your agent workpoint its MCP at agentklar mcp

# the agent claims, writes code, and submits — over MCP
echo '{"jsonrpc":"2.0",...,"method":"claim_task",...}' | ./agentklar mcp

4Gate, then approveonly you finish it

./agentklar gate AK-1        # runs your checks, keeps the proof
./agentklar approve AK-1     # your call — the agent can't make it

That's the whole loop. Want the state machine, evidence model, and the community library? See all features →