A human defines acceptance criteria and a verification method before an agent can claim work.
Product / The operating model
Coordination that outlives the conversation.
Agentklar is the local control plane between a team, its coding agents, and the repositories they change. It turns scattered agent sessions into scoped work with durable context and reviewable proof.
01 / Workflow
A lifecycle that cannot be skipped.
One holder wins an atomic claim. Leases and fencing reject stale workers.
The exact submission is frozen for review before quality checks run.
Declared recipes run and retain machine-attested evidence.
A person reviews the task and its evidence snapshot.
Only a nonce-bound human approval crosses the final boundary.
02 / Control center
One view across every registered project.
The catalog knows where projects live and which workspace belongs to each one. Protected task state remains in the project's own store, avoiding global ID collisions and risky migrations.
Overview and attention
See project activity, blocked work, pending approvals, and operational alerts without launching a board per repository.
A board that respects state
Create, edit, filter, prioritize, assign, reorder, and link dependencies. Moves call the workflow engine; dragging never writes a fake state.
Project-bound MCP
Agents claim only from the repository where they started. Cross-project browsing is for the human interface, not automatic context mixing.
Evidence before approval
Review the command, working directory, exit code, timing, and retained log hash for the exact submitted snapshot.
03 / Intelligence
Three kinds of continuity. No hidden state.
Agentklar separates durable instructions from generated learnings and derived retrieval. People can inspect every layer and see where it came from.
.agentklar/knowledge/Knowledge
Checked-in architecture, decisions, conventions, glossary, and runbook. Versioned and human-reviewable.
memory.sqliteMemory
Cross-session learnings with project, task, holder, source, and time provenance. Humans control deletion.
context.sqliteContext
A searchable index that retrieves focused excerpts for a task instead of loading every document into every prompt.
04 / Trust boundary
The agent cannot grade its own work.
The agent-facing tool surface stops at requesting an approval presentation. It has no operation that can approve, reject, or mark a task Done.
claim_taskAGENT
submit_for_reviewAGENT
record evidence / QASYSTEM
approve / reject / DoneHUMAN ONLY
05 / Ownership
Local by design. Loopback only.
One Go binary embeds the UI and uses SQLite for durable state. No frontend runtime, hosted account, or external tracker is required.
Run on loopback
agentklar ui --open starts the control center on your machine and binds it to 127.0.0.1.
Not a network service
The control center does not listen remotely. Shared-network deployment and team identity are outside this release.
Vikunja is optional legacy integration. Existing users can keep its projected board and reconciliation flow. New setups use Agentklar's native control center.