A control room for coding agents
Kandev starts from a problem that appears only after coding agents become genuinely useful. One Claude Code or Codex session is easy to watch in a terminal. Five parallel tasks across several repositories are not. Prompts, branches, worktrees, review notes, and pull requests scatter across windows, while the human loses track of what is ready and what needs intervention. Kandev turns that work into tasks on a board and gives each task an isolated execution and review space.
The supported agent list is unusually wide. It includes Claude Code, Codex, GitHub Copilot, Gemini CLI, Amp, OpenCode, Cursor, Devin, Qwen, and many others. Most structured integrations communicate through Agent Client Protocol adapters, while arbitrary terminal agents can run through a custom TUI profile.
That neutrality is its clearest advantage. A workflow can use one profile for planning, another for implementation, and a third for review. Tasks can run concurrently in separate Git worktrees, preventing two agents from editing the same checkout. Multi-repository tasks get a worktree, branch, pull request, and changes grouping for each repository.
More than a kanban wrapper
The board is only the front door. Open a task and Kandev provides agent chat, a terminal, file tree, editor with language-server support, Git changes, an embedded browser preview, and review controls. Sessions can be resumed and prompts queued. Workflows are exportable as YAML, making a review process portable between workspaces or installations.
Agents may run as local processes, in Docker containers, on servers through SSH, or in Sprites cloud environments. Profiles hold preparation scripts, environment variables, credentials, custom prompts, and executor settings. Kandev also exposes an external MCP endpoint so another compatible agent can manage tasks, plus session-scoped MCP tools that let an agent create subtasks, message related work, and inspect conversations.
There is a notable gap for custom TUI agents. An open August 9 issue explains that these profiles do not receive the session-scoped MCP configuration, and the interface offers no way to enable it. The wrapped CLI still runs, but it cannot call Kandev's task coordination tools. Users bringing an unusual agent should distinguish plain terminal passthrough from full participation in Kandev's orchestration model.
Office mode should also be treated carefully. The README calls it an in-progress, feature-flagged autonomy layer and says it will be documented as supported only after it is live. Roles, routines, approvals, memory, budgets, and persistent teams describe the direction, not a finished reason to adopt the product today.
Quick install, substantial onboarding
Homebrew on macOS or Linux and Scoop on Windows install native bundles without requiring Node.js. The npx route needs npm 7 or newer, and a global npm package is also available. Kandev stores its SQLite database and worktrees under ~/.kandev by default. Starting the interface is therefore simple. Configuring a dependable working environment is the larger job.
Each agent still needs its own credentials, subscription, models, and executable behavior. Repositories need safe secret handling and a known branching strategy. Docker and SSH executors add image, network, host, and credential decisions. Kandev organizes those costs but does not remove them.
Runtime repeatability needs scrutiny. The README says Kandev does not pin managed npm runtimes for Claude, Codex, OpenCode, Copilot, or Gemini. Normal launches can use npm's best-effort cache, and updates can be requested explicitly from settings. A current issue goes further, reporting that unversioned npx resolution puts npm registry access on the launch path for managed agents and does not record the resolved bridge version. That is a poor fit for offline, regulated, or strictly reproducible environments.
Self-hosting has sharp edges
Kandev is designed as a server that can be reached from another device. The README recommends Tailscale or another VPN, and the roadmap still lists secure access controls under current work. Follow that boundary. A system that holds repository access, agent credentials, terminals, and code changes should not be placed directly on the internet merely because the web interface loads.
The service tooling is detailed on Linux and macOS. It supports user-level systemd or launchd units and system-wide installations, documents boot behavior, health checks, logs, update channels, and recovery. Windows service mode is not supported. More importantly, version 0.86.0 explicitly labels the macOS and Windows desktop installers as unsigned development builds. An open Windows report describes Microsoft Defender quarantining the bundled agentctl.exe, after which agent launches failed without a clear explanation.
AGPL 3.0 is another deliberate choice. It is friendly to users who want the whole server to remain open, but companies modifying and providing it over a network should have counsel review their obligations. This is not the permissive-license default common among small developer tools.
Health and the decision
Kandev was created in January 2026, pushed on August 10, and released version 0.86.0 the same day. Nine other listed releases arrived between July 13 and August 4. Current pull requests cover testing, settings, Kanban filtering, plugins, executors, and source-control integrations. GitHub's 32 open count combines issues and pull requests, and the majority of the recently updated items are active pull requests. The project is clearly alive.
The velocity is encouraging and destabilizing at once. Release 0.86.0 contains a very long list of features, fixes, refactors, translations, and dependency changes. That pace can close gaps quickly, but teams should pin a tested version, read release notes, and rehearse upgrades instead of assuming every weekly build is interchangeable.
Kandev is worth trying when coordinating agents has become harder than launching them. Its integrated review surface and worktree model are more valuable than the board's visual appeal. For a single developer with one favorite agent, Cline or the native CLI is simpler. For autonomous platform work, compare OpenHands. Kandev earns its place when provider choice, parallel tasks, and human review gates all matter enough to justify running another control plane.