Koda is a desktop command deck where AI agents do the building. Run Claude, Codex and more in tiled panes, hand a goal to Autopilot, and let them coordinate on a shared board — while your AI Second Brain briefs every agent on who you are.
Windows · macOS · Linux — 7-day free trial · no credit card required
Koda gives the agent CLIs a real home — organized, persistent, and designed for running many of them at once.
Run a dozen agents at once and they won't trip over each other. Every session announces itself on a shared board, claims a lane, and keeps its card up to date — live, with nothing to wire up. Included in your 7-day free trial — unlimited agents on Pro.
Every session's first prompt drops a card on the board — titled, tagged, automatic. No one has to remember.
Before touching code, each agent calls koda_board and koda_roster — then takes a slice that won't collide.
Each agent claims its card and declares a lane like backend · src/api — two agents never own the same work.
Cards move themselves Backlog → In Progress → Review → Done. The board is always a live picture of the whole team.
Koda ships a tiny MCP server into each workspace and registers it for every agent. They coordinate by calling tools; Koda holds the state on your machine.
The same bus carries KodaMemory — durable project knowledge across sessions. Works with Claude Code, Codex, Gemini & Grok. State lives in a gitignored .koda/coord/ — local-first, no backend. Incoming messages gently wake the idle pane.
Autopilot turns a goal into shipped code. A lead agent plans the work on the board, builds tests-first, and loops until your verify command passes. Turn on parallel workers and each card builds in its own isolated git worktree, merged back when green.
Your task board becomes an execution engine. Break work into cards, assign each an agent, and launch them in parallel — you plan and review, the agents build.
Open the full diff of everything an agent touched, measured against a snapshot baseline. Comment on lines, send fixes back to the same session — nothing ships without your sign-off.
KodaMemory keeps what your agents learn. Every fix, decision and gotcha becomes a node in a local knowledge graph — and each new session recalls exactly what's relevant to its task. No lesson gets relearned twice.
.koda/coord/ — no backend, nothing syncedKodaMemory remembers your project. Your Second Brain remembers you — and it briefs every agent, in every workspace, automatically.
A short guided interview seeds it; from then on it learns ambiently from your sessions — even how you write, so agents mirror your voice. Every prompt gets briefed with just the facts that matter. And the Reading Queue turns saved links into knowledge: paste a URL, Koda reads it and banks the learnings.
KodaVoice is push-to-talk voice typing for Windows. Hold a hotkey in any app — agent panes included — speak, release, and your words are typed at the cursor. Whisper runs entirely on your machine: private, offline, nothing uploaded.
No proxy, no lock-in. Koda launches the CLIs you already have installed.
Point Koda at any project. It becomes a workspace with your file tree, editor, browser and task board — your layout saved for next time.
Spin up tiled terminals and launch Claude, Codex, or your own custom agent profile into each one. Keys from the vault are injected automatically.
Drop tasks on the board, hit Run, and watch agents build in parallel. Review their work in the editor and preview it live in the browser.
Koda runs the real first-party CLIs in real terminals. It never proxies your API, never extracts or reuses your OAuth tokens, and never sits between you and your provider. You sign in to your tools exactly as you always have.
That means you stay fully inside every provider's terms — and your credentials never leave your device.
The UI, your terminals and the editor all re-theme together — live.
Full access from day one — no credit card required. Bring your own agent subscriptions; Koda is the workspace, you keep your logins.
Koda is a desktop development environment built around coding agents. Instead of writing every line yourself, you orchestrate AI agents — Claude, Codex and others — across tiled terminal panes, launch their work from a Kanban board, and review the results in a built-in editor and browser. It's an ADE — an agentic development environment — designed for building software with agents.
You bring your own agent credentials. Koda launches the agent CLIs you already use, so you sign in with your existing subscription or drop an API key into the encrypted vault. Koda never proxies your account — it just gives those tools a great home. As for Koda itself: create an account and every feature is free for 7 days, no credit card required. After the trial it's $15/mo (or $144/yr) to keep going.
Any CLI-based agent. Out of the box that includes Claude Code, OpenAI Codex, Gemini, Grok and opencode, and you can define your own custom agent profile with a name, launch command and provider in a few seconds.
No. Koda is local-first: your code, shells and agents run on your own machine. API keys are encrypted by your operating system and decrypted only inside Koda's main process — never exposed to the interface or sent to us.
Koda is a desktop app built on Electron, with installers for Windows, macOS (Apple Silicon & Intel) and Linux (.AppImage / .deb). The Windows build auto-updates itself in the background. You can grab the latest build for any platform from the releases page.
Because Koda isn't code-signed with a paid platform certificate yet. Every operating system shows a one-time caution the first time you open any app that isn't signed — it's about the certificate, not the app, and it's completely expected. Koda is local-first and fully open about what it does: your code, shells and agents stay on your machine. Here's how to get past the prompt, once, per platform:
macOS — the .dmg installs normally; on first launch, right-click (or Control-click) the Koda app and choose Open, then Open again in the dialog. After that it launches like any other app. (If you double-clicked and got "Apple cannot check it for malicious software," just close that dialog and use right-click → Open instead — or go to System Settings → Privacy & Security and click Open Anyway.)
Windows — if Microsoft Defender SmartScreen appears, click More info, then Run anyway. The installer proceeds normally and future updates apply in the background.
Linux — mark the AppImage executable (chmod +x Koda-*.AppImage, or Properties → Permissions → "Allow executing file as program") and run it. The .deb installs via apt or your package manager as usual.
Signed and notarized builds are on the roadmap — once they land, these prompts disappear and macOS in-app auto-update switches on automatically.
Yes — that's the point. Tile up to 8 agents side by side, each in its own pane, and launch them in parallel from the task board. Per-pane account switching even lets different panes use different logins.
Autopilot is Koda's hands-off mode Pro. You describe a goal and an acceptance check (and optionally a verify command like npm test), and a lead agent plans the work on the board, builds it tests-first, and loops until your verify command passes — Koda keeps the session moving so you don't have to babysit it. Turn on parallel workers and the lead splits the goal into independent cards; Koda runs a worker per card in its own isolated git worktree and merges each back when its tests are green. Run it supervised (it asks before each action) or in auto mode, and it pauses to ask only when it hits a decision that's genuinely yours to make.
Koda runs a small MCP coordination server in every workspace and registers it for each agent. The moment a session starts it auto-creates a card on the shared Kanban board (titled with its task, tagged with its pane) and keeps that card updated as it works. Before editing, every agent reads the board and roster to see who's doing what, then claims a lane so two agents never own the same work. Coordination is included in the 7-day free trial and stays unlimited on Pro. It auto-enables for each workspace you open (except ones you mark Sensitive), and you can flip it off any time with the Coordination toggle in the Tasks tab — there's nothing else to wire up. The shared state lives in a gitignored .koda/coord/ folder on your machine; nothing is proxied and there's no Koda backend.
KodaMemory is a shared knowledge graph Pro that lets your agents remember what they discover. As agents work, they record durable knowledge — a decision, a gotcha, a where-things-live pointer — as nodes, and link them with relationships like caused_by or supersedes. When a new session starts, Koda ranks that graph against the task at hand and injects the most relevant memory — not just the last few notes — so knowledge compounds instead of getting relearned. It's deduped and compacted automatically, and like everything else it lives locally in .koda/coord/ with no backend. See the KodaMemory docs for the full model.
Where KodaMemory remembers your project, the Second Brain Pro remembers you: who you are, what you're building, your goals, preferences, skills and how you like answers written. You seed it with a short guided interview, and after that it learns from your sessions automatically — facts you stated outright are saved silently, while anything Koda merely inferred waits in a review digest for your one-click Keep or Dismiss. When you prompt any agent in any workspace, Koda injects just the facts relevant to that task, so outputs come back tailored to you instead of generic. Everything lives locally in ~/.koda/brain — outside every repo, never committed, never uploaded — and every fact can be edited or forgotten in one click from the "You" panel. See the Second Brain docs.
Download Koda and turn your desktop into a command center for coding agents.
Need a different build? See all platforms on the releases page →