KaibanJS

A KaibanJS program is a Team of agents and tasks. Each task names the agent that performs it and an expected output. The docs describe three deterministic execution modes that can be combined: tasks run in list order, start once their dependencies finish, or run side by side when marked allowParallelExecution. An LLM manager that would re-plan tasks at runtime is listed as coming soon, so the order is always defined by the developer.

Agents come in three types. The default ReactChampionAgent runs an LLM reasoning loop with tools (LangChain-compatible tools are accepted). WorkflowDrivenAgent executes a step graph from @kaibanjs/workflow, and ExternalCodingAgent hands a task to a local Claude Code or OpenCode process, one CLI call per task.

All state sits in a Team Store, described as Redux-inspired, which records task statuses (TODO, DOING, BLOCKED, REVISE, AWAITING_VALIDATION, DONE) and workflow logs; UIs subscribe to it. Team-level memory is on by default, so every task sees earlier results; with memory: false, tasks only receive results referenced through {taskResult:taskN}.

A task created with externalValidationRequired: true moves to AWAITING_VALIDATION when its agent finishes, and the workflow status becomes BLOCKED. The application listens with team.onWorkflowStatusChange(), lists waiting tasks with getTasksByStatus('AWAITING_VALIDATION'), and either approves with validateTask(taskId) or sends provideFeedback(taskId, text). Feedback moves the task to REVISE, and the agent reworks it; feedback can also be given on tasks that are BLOCKED or already DONE. Each task keeps a feedbackHistory with content, status and timestamp. ExternalCodingAgent appends reviewer feedback to the prompt before calling the CLI again.

npx kaibanjs@latest init

import { Agent, Task, Team } from 'kaibanjs';

const researcher = new Agent({ name: 'Ada', role: 'Researcher', goal: 'Collect facts on a topic' });
const writer = new Agent({ name: 'Ben', role: 'Writer', goal: 'Turn research into a short brief' });

const research = new Task({ referenceId: 'research', description: 'List key facts about {topic}',
  expectedOutput: 'Bullet list', agent: researcher });
const brief = new Task({ description: 'Write a 100-word brief from: {taskResult:task1}',
  expectedOutput: 'Short brief', agent: writer, dependencies: ['research'], externalValidationRequired: true });

const team = new Team({ name: 'Brief team', agents: [researcher, writer], tasks: [research, brief],
  inputs: { topic: 'solid-state batteries' }, env: { OPENAI_API_KEY: process.env.OPENAI_API_KEY } });

team.onWorkflowStatusChange((status) => {
  if (status === 'BLOCKED') team.getTasksByStatus('AWAITING_VALIDATION').forEach((t) => team.validateTask(t.id));
});
const output = await team.start();
console.log(output.status, output.result);

• npx kaibanjs@latest init scaffolds a Kaiban Board project; the key goes in .env as VITE_OPENAI_API_KEY (or another provider), then npm run kaiban restarts the board. For library use, install with npm install kaibanjs.
• The quick start lists Node.js 14+ and npm 6+ as prerequisites.
• Anonymous telemetry is on by default; set KAIBAN_TELEMETRY_OPT_OUT=true to disable it.
• ExternalCodingAgent needs Node.js (it spawns child processes), the claude or opencode CLI on PATH, and has a default timeout of 10 minutes per call.
• A2A and OpenClaw adapters must run server-side, not in browser bundles.

Partly. Since 0.20.1 its agents can use MCP tools through LangChain's @langchain/mcp-adapters client over stdio or SSE. KaibanJS does not ship its own MCP module or server.

The repository's last push was on 2026-05-15 and the newest release is v0.24.1 from 2026-05-01, so there has been no activity for several months.

The library is MIT-licensed. Its maker, Kaiban, also runs the separate KaibanIO platform, which the docs connect to over A2A; no public pricing page for it was found.

Yes. ExternalCodingAgent runs a task through the Claude Code or OpenCode CLI, and an OpenClaw plugin exposes a whole team as a single tool.

Mark a task with externalValidationRequired; when it reaches AWAITING_VALIDATION, call validateTask() to approve or provideFeedback() to send it back for revision.