Kangentic
Drag a card and an AI coding agent starts — every session on one board that flags the ones waiting on you.
- Source repo
- Kangentic/kangentic
- Stars
- ★ 140
- Last updated
- today
- License
- AGPL-3.0
- Primary language
- TypeScript
- FA score
- 52/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeClaude.ai (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Solo developers running several Claude Code or Codex sessions who lose track of which agent is waiting on a prompt.
- Not a fit if
- Developers who prefer one task start-to-finish over parallel agents
- Teams that want agents running in the cloud instead of on their machine
- Users who rely on Cursor or Claude Desktop rather than command-line agent CLIs
- Source review
- 52/100 · Major gaps
What does this agent do, and when should you use it?
Kangentic is an Electron, React and TypeScript desktop 'agentic development environment' that orchestrates multiple AI coding CLIs from a single kanban board. Each task is a card showing its model, context-window usage, and the agent's latest message, with a marker when it needs your input. Dragging a card into a column creates a git worktree, applies that column's permission mode, launches the configured CLI, and streams it in a live terminal below the board. Columns carry their own agent, model, effort and permission mode, and the board setup is committed with the repo in kangentic.json while tasks and sessions stay in a local SQLite database. A bundled MCP server lets a running agent create tasks, move its own card, search past conversations or message another task's agent; Automations, Workflows and the Changes panel tie multi-step pipelines, diffs and pull requests together. Fifteen agent CLIs are supported, from Claude Code and Codex to Aider, OpenCode, Goose and Ollama.
It reads your repositories plus kangentic.json and renders one board per project. Tasks are created with a title and prompt, or imported from GitHub Issues, GitHub Projects, Azure DevOps or Asana. When a card enters a column, Kangentic creates a git worktree, applies the column's permission mode, and starts the matching agent CLI — Claude Code, Codex CLI, Gemini CLI, Cursor CLI, GitHub Copilot CLI, Grok Build, Antigravity CLI, Goose CLI, OpenCode, Aider, Qwen Code, Kimi Code, Factory Droid, Warp (Oz CLI), or Ollama for single-turn local chat. During the run the card shows model, context usage and the latest agent message, and the live terminal opens below for answering prompts. The Changes panel shows the task's diff, commit history and the merge readiness of a linked GitHub or Azure DevOps pull request. Automations fire when a task enters a column, when its agent finishes a turn or when the task leaves, sending a message or slash command, running a script, calling a webhook, notifying you, or moving the task onward, with plain-language conditions an agent evaluates. Workflows let a workflow agent plan linked tasks from a brief and start each as its dependencies reach Done, across multiple projects. The MCP server exposes board tools; data lives in local SQLite; Kangentic Mobile (end-to-end encrypted, via Kangentic's relay or your own) extends monitoring and replies to your phone.
- Solo developers running several Claude Code or Codex sessions who lose track of which agent is waiting on a prompt.
- Contractors juggling tasks across multiple clients or repos who want one Agent Monitor view over all running agents.
- Teams already paying for Claude Code, Codex and similar CLIs who want to mix them in one pipeline.
- Engineering teams with multi-step plan → execute → review → merge flows that need per-project agents and automations.
- Developers who want to answer an agent or review a diff from their phone after stepping away from the desk.
- Teams whose work arrives as GitHub Issues, Azure DevOps or Asana tickets and should become runnable tasks directly.
How do you install or deploy this agent?
Install the latest platform build with npx; the app updates itself afterwards:
npx kangenticInstallers are also available on the GitHub Releases page if you prefer to run one yourself.
Requirements:
- Windows x64, macOS 13 or later on Apple Silicon or Intel, or Linux x64/arm64 (.deb or .rpm). The Knowledge Graph is unavailable on Intel Macs.
- Node.js 20 or later for npx, and Git 2.26 or later.
- At least one supported agent CLI installed and signed in (for example Claude Code or Codex CLI).
Building from source needs Node.js 22.14 or later, or 24 or later:
git clone https://github.com/Kangentic/kangentic.git
cd kangentic
npm install
npm startWSL users install the Windows version and can pick a WSL distro as the terminal shell; agents run the Windows install of each CLI, so a CLI installed only inside the distro is not found.
How do you use this agent?
On first launch, shape the board: keep the default columns or rename, add and rewire them, and pick the agent, model, effort and permission mode each column runs. That setup is written to kangentic.json at the repo root and shared with teammates through git. Then create a task with a title and prompt, or import one from GitHub Issues, GitHub Projects, Azure DevOps or Asana. Dragging the card into a column runs it:
To Do → Planning → Executing → Code ReviewKangentic creates the worktree, applies the column's permission mode and starts the agent; follow it in the terminal below the board, answer when it asks, and drag the card on when it is done. Use the Changes panel to review the diff, commit history and linked pull-request state. For mobile access, pair your phone once with Kangentic Mobile (App Store or Google Play); the connection is end-to-end encrypted through Kangentic's relay or one you host. To opt out of the default telemetry, set the environment variables before launching:
export KANGENTIC_TELEMETRY=0
export KANGENTIC_ERROR_REPORTING=0What are this agent's strengths and limitations?
- Board-driven execution: dragging a card creates the git worktree, applies column-level permissions and launches the agent, moving orchestration cost from you to the card.
- Wraps fifteen agent CLIs (Claude Code, Codex, Gemini, Copilot, Aider, OpenCode and more) using credentials you already configured — no API key, proxy or Kangentic account.
- Column layout is committed in kangentic.json so teammates share the same workflow, while tasks and sessions stay on each person's machine.
- Bundled MCP server lets a running agent create tasks, move its own card, search past conversations and message another task's agent.
- The Changes panel consolidates a task's diff, commit history and GitHub/Azure DevOps pull-request merge status in one place.
- Kangentic Mobile is end-to-end encrypted and works through the official relay or a self-hosted one, so you can answer agents away from the desk.
- Runs only on Windows x64, macOS 13+ and Linux x64/arm64 via .deb or .rpm; the Knowledge Graph is unavailable on Intel Macs.
- Depends on agent CLIs already installed and signed in on the host; under WSL, only the Windows install of each CLI is detected.
- Release builds send anonymous usage events (Aptabase) and error reports (Sentry) by default, with no in-app switch — only environment variables disable them.
- Interface translations are machine-made and open to review, so non-English wording may need correction.
- Building from source requires Node.js 22.14+ or 24+, and contributors must sign a CLA before a first pull request is merged.
- Data lives in local SQLite; there is no documented multi-user or multi-device sync beyond the board configuration file.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Kangentic This agent | 52 · Major gaps | Desktop appFree + model costs | ★ 140 | today | TypeScript | Codex · Claude Code |
| wmux Agent Workspace | 74 · Some gaps | Desktop app | ★ 447 | today | TypeScript | Codex · Claude Code |
| Ateam Agent Workbench | 70 · Some gaps | Desktop appFree + model costs | ★ 16 | today | TypeScript | Codex · Claude Code |
| AgentsMesh | 45 · Major gaps | Self-hosted serviceFreemium | ★ 2.4k | 1mo ago | Go | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows an Electron app with all fuses explicitly set (RunAsNode/NodeOptions/inspection disabled, ASAR integrity on), which is a plus; however agents run with the user's existing CLI credentials and can run scripts, call webhooks and use a mobile relay, and no fine-grained least-privilege model is documented, so least_privilege is 1. User confirmation: task deletion can skip confirmation, kangentic.json silently overrides the DB on project open, and stopScript/Run script automations execute on teammates' machines without prompt, so 1. Data-flow transparency: README states local SQLite storage and default-on Aptabase/Sentry with env-var opt-out, but no in-app switch, so 2. Sensitive data: only claims task text is excluded from telemetry; no credential/secret storage policy, so 1. Dependency security: packaging doc details native-module whitelist, afterPack verification and ONNXRUNTIME install-script skipping, so 2. External effects: automations can fire webhooks, run scripts and move cards with no default sandbox or rollback, so 1. Rollback: only auto-update and uninstall are covered; no undo for agent changes or automations, so 1. Source attribution: AGPL and CLA present, but publisher identity is unverified, so 1.
Self-consistency: README, installation guide and packaging doc agree on platforms, dependencies and update paths, so 2. Dependency availability: Node/Git versions, per-platform installers and system libraries are spelled out, so 2. Failure messages: installation doc gives concrete troubleshooting for SmartScreen, Gatekeeper, icacls and native-module builds, so 2.
Audience and scenarios: README explicitly lists who it fits and who it does not, so 2. Capability boundaries: documents Intel Mac Knowledge Graph gap, WSL limits and per-CLI support differences, so 2. Trigger precision: automation conditions are judged by an agent, and file watching coexists with silent apply-on-open, making trigger semantics complex and error-prone, so 1. Environment fit: covers Windows/macOS/Linux including arm64, so 2.
Information architecture: docs split into platforms, settings and user guide, clearly structured, so 2. Install notes: very thorough, covering npx, manual packages, source build and uninstall, so 3. Naming stability: docs admit several config keys and automation types were renamed/retired with migrations, so 1. Examples and FAQ: config examples and troubleshooting sections exist but no dedicated FAQ, so 2. Known limitations: limitations are stated in several places, so 2. License: full AGPL-3.0 text present, so 3. Versioning/changelog: only badges and release-page links; no in-repo CHANGELOG, so 1. Maintenance responsibility: publisher unverified, only email and Discussions, so 1.
Output usability: board, diff and terminal outputs are practical for multi-agent workflows, so 2. Marginal value: addresses the real pain of tracking parallel agents, so 2. Cost-benefit: requires a desktop app, at least one signed-in CLI, default telemetry and a broad permission surface; payoff depends on heavy multi-agent use, so 1.
Claim traceability: docs cite sources pointing at specific files, but only docs were supplied, not the source, so 1. Cross-source corroboration: README and installation/packaging docs corroborate some facts, but core behavior has no test evidence, so 1. Fact/inference separation: docs mix design intent and inferential prose without verifiable markers, so 1.
- Agents run with the user's existing CLI credentials and can execute scripts, call webhooks and be driven remotely via the mobile relay; the permission surface is broad, so use in an isolated environment or dedicated account.
- kangentic.json silently overrides the database on project open, and stopScript/Run script automations execute on teammates' machines without prompt; review shared configs before committing.
- Telemetry and error reporting are on by default with no in-app switch, only environment variables.
- Publisher identity is unverified; maintenance responsibility and update-path reliability cannot be confirmed from the repository.
- This is a static documentation review; no code was executed and behavioral claims were not independently verified.