Self-Hosted AI Agent Kanban Board
KittyClaw is a local control plane for software work performed by AI agents. Start with a ticket, follow progress on the board, and read the run that changed the code. Configure automatic checks and production delivery, or keep a human release checkpoint. Project data is stored locally; agent context is sent to the provider you select.
Get started on GitHub AGPL-3.0-or-later · single binary · SQLite per project
What is an AI agent kanban board?
A traditional kanban board tracks work that humans do. An AI agent kanban board goes further: a configured workflow can start agents, check their output, request corrections, and continue to delivery. The board shows what is waiting, running, checked, or paused at a chosen human checkpoint. Test results and delivery evidence stay linked to the ticket so you can check the result.
KittyClaw connects three parts of that journey: a live board, verifiable work, and configurable delivery. Define acceptance criteria up front and configure passing and failure routes. Your pipeline can run tests and checks through to production deployment, with human review only at the steps you choose.
Why self-hosted matters for AI agents
When an agent has permission to read your repo, run your tests, and open pull requests, where that agent runs is a security and IP question, not a convenience one. KittyClaw runs as a single self-hosted binary with a local SQLite database per project. Your tickets, your code context, and your pipeline configuration stay on your infrastructure.
- Local-first: one process, one SQLite file per project. No account, no cloud tenant.
- Choose the agent backend: route work to Claude Code, OpenAI Codex CLI, Grok Build, Mistral Vibe, or local models through Ollama. Per-ticket token cost tracking (introduced in v0.10) shows what each run spends.
- AGPL-3.0-or-later: inspect, modify, and self-host the source under an open-source license.
Open-source alternative to Linear, Jira & GitHub Projects
Linear, Jira, and GitHub Projects have all shipped AI agents in 2025–2026. They are excellent tools. But if your priority is owning the orchestration layer — self-hosting it, choosing the model, and giving each processing stage a mission and persistent memory — the trade-offs differ. Here is an honest comparison, current as of July 2026.
| Capability | KittyClaw | Linear | Jira (Rovo) | GitHub Projects (Copilot) |
|---|---|---|---|---|
| License | AGPL-3.0-or-later, open source | Proprietary | Proprietary | Proprietary |
| Self-hosted / local-first | Yes — single binary, SQLite | No (SaaS only) | Data Center (enterprise, quote-based); native AI needs a Cloud plan | Cloud agent (GitHub-hosted runners) |
| Data stays on your infra | Yes | No | Cloud required for AI features | No |
| AI agents assignable to work items | Yes — as full members | Yes (delegate; human stays primary assignee) | Yes (Rovo agents) | Yes (assign an issue to Copilot) |
| Agent ships code autonomously | Yes (programmer → commit/PR) | Via connected third-party agents | Yes (Rovo Dev) | Yes (opens a pull request) |
| Stage missions + processor memory | Yes (reusable skills + persistent processor lessons) | No | Build agents in Rovo Studio | Custom agents (cloud-managed) |
| Pipeline-based dispatch engine | Yes (named pipelines, column processors, routing) | Workflow rules | Automation rules | GitHub Actions |
| Model choice | Claude, Codex, Grok, Mistral + local Ollama | Managed | Managed (Rovo credits) | Managed (Copilot) |
| Price of the app | $0 self-host (you bring the model key) | $10–$16 / user / mo | Rovo Dev $20 / dev / mo + credits | Copilot Pro from $10 / mo |
Prices and features re-verified from primary sources on 2026-07-27. Competitor products evolve fast; figures are point-in-time.
How to read this table
- Linear treats agents as delegates: you connect an external agent, delegate an issue, and the human remains the primary assignee while the agent contributes. Great interoperability, no self-hosting.
- Jira Rovo is the deepest enterprise suite — agents can triage, route, and (via Rovo Dev) write code. The catch for self-hosters: the native AI features require an Atlassian Cloud plan; Data Center syncs to the cloud rather than running AI locally.
- GitHub Projects + Copilot is the smoothest path if you already live in GitHub: assign an issue to Copilot and it spins up a sandbox, runs your tests, and opens a draft PR — on GitHub-hosted runners.
- KittyClaw keeps the board, agent runs, and verification history on your machine. Configure stage missions, automatic verification, and delivery steps. Add human checkpoints where you choose.
Use your coding tools in a workflow you can review
Choose Claude Code, OpenAI Codex CLI, Grok Build, or Mistral Vibe for agent runs. DeepSeek and local Ollama models use Claude Code as their transport. Give each workflow step a mission, a model, and reusable project skills.
- Define the outcome: describe the behavior you want and the acceptance criteria.
- Follow the work: read the run, inspect changes and test results, and steer or stop an active agent.
- Validate and deliver: configure independent checks, corrections, and delivery routes, with optional human checkpoints.
Processor memory preserves lessons across runs. Optional JEV judgments can help with triage, profile selection, and evidence review. Configure owner decisions and approval requirements for the steps where human authorization is needed.
Start with a small, verifiable change
Install the Windows app with its included .NET runtime, or run from source on Windows, macOS, or Linux with the .NET 10 SDK. Install and authenticate an agent provider separately. Choose a workspace, review the workflow proposed during setup, and send a small ticket through it. Check the first run and refine the criteria and routes before extending the automation. KittyClaw is open source under AGPL-3.0-or-later; your model provider and optional services have their own pricing.
FAQ
Is KittyClaw really self-hosted?+
Yes. The app and project databases run locally. Agent context goes to your selected model provider, and optional JEV reviews use TypeSafe. Production builds send an anonymous daily instance heartbeat.
Is it open source?+
Yes — v0.12 and later are licensed under AGPL-3.0-or-later. The source is on GitHub for you to inspect, modify, and self-host under that license.
Which AI backends does it support?+
KittyClaw supports Claude Code, OpenAI Codex CLI, Grok Build, Mistral Vibe, and local models through Ollama, with per-action and per-member routing. Per-ticket token cost has been visible on the board since v0.10.
How is this different from Jira Rovo or GitHub Copilot's coding agent?+
KittyClaw keeps project workflow and run history locally and supports several coding providers. Configure stage missions, automatic checks, and delivery up to production, with optional human release decisions.
What does it cost?+
The app is free to self-host ($0). You pay only your own model provider (e.g. Claude API usage), or run local models with Ollama at no per-token cost.