Koda
kodahq.io- Category
- AI Agents
- Rank
- No. 2149Tools index
Previous survey · No. 2013 ·
- Pricing
- Free
- Type
- APP
- Use case
- Coding
- Interfaces
- Desktop
- Builder
- gr0x01
- GitHub
- 1 stars
- Latest release
- v0.1.21
- Date
About
Koda is a free Mac app that acts as an AI coding agent, letting you build software using your existing Claude or ChatGPT subscription instead of buying API credits. It manages a real project folder with file previews, guardrails, versioned save points, and step-by-step approval controls so you can review or undo agent actions as it writes and runs code.
What it does
Koda is an Electron desktop shell for macOS that wraps the real Claude Code and OpenAI Codex command-line agents rather than calling a model directly. You talk to the agent in a window; Koda starts the bundled engine, signs you in with your existing plan, shows each tool request for approval, and records its own undo history beside your git. A bundled pack of rules, skills and subagents shapes the agent's behavior, and each rule can be switched off or reworded per project.
Why it's ranked here
The case rests on engineering care in the places that usually break. Sign-in detection runs under the same environment the engine will use and strips a stray Anthropic key so it cannot quietly bill at API rates. Stored keys are encrypted through the macOS Keychain and never reach the interface. File reads check that the opened file is the same one that passed the project-boundary check, closing a symlink swap race. The licence is AGPL-3.0. The limits are real too: Mac only, early version numbers, and continuous integration that lints and builds but does not run the test suite.
What's good
The two engines are described by one capability table instead of scattered checks, so it is explicit that Claude absorbs a message sent mid-turn while Codex queues it, and that Codex has no native plan mode, so Koda's own approval gate enforces it. The renderer runs sandboxed with context isolation and a strict content security policy in packaged builds. Saving a rule or skill takes a checkpoint first, so a bad paste is recoverable. Voice dictation is transcribed on the device.
Tradeoffs
It is macOS only: the packaging targets a disk image and zip for Mac, and the voice and assist helpers are Mac-specific. Continuous integration runs lint and a full build on Linux, while unit and end-to-end tests stay local. Sync, backup and phone control live in a separate private repository as a paid tier; the open client ships an inert stub for remote control. The trademark notice forbids shipping a fork under the Koda name. The automatic switch to an API key when a plan limit hits works only on the Anthropic side.
How to use it well
Pick it if you already pay for Claude or ChatGPT and want those agents in a desktop window with approvals and a separate undo trail, especially for people who would not live in a terminal. Keep your own git history; Koda's checkpoints are deliberately separate from it. Review the bundled rules in settings before the first long session and switch off what does not fit. It does not cover Linux or Windows, and it adds no model of its own: output quality is whatever the underlying engine delivers.
Technical notes+
src/main/auth.ts drives claude auth login --claudeai over piped stdio, parses the OAuth URL from stdout, writes the pasted code to stdin, and times out after five minutes; classify() flags an apiKeySource as a billing trap unless the key was injected deliberately. src/main/api-key.ts stores per-engine keys via Electron safeStorage as base64 ciphertext with file mode 0o600 and fails soft to no key. shared/engine-capabilities.ts declares CLAUDE_CAPABILITIES (native plan mode, broker approvals via MCP, apiKeyFallback true) against CODEX_CAPABILITIES (turnText plan mode, native approvals). src/main/contained-read.ts opens with O_NONBLOCK, compares dev and inode against the realpath, and rechecks size, mtime and ctime after reading. .github/workflows/ci.yml runs only lint and build.
Observed
- License
- AGPL-3.0 (LICENSE), with a separate trademark restriction on the Koda name (README.md)
- Language
- TypeScript, Electron with React (package.json)
- Platform
- macOS only; packaged as dmg and zip (electron-builder.yml)
- Agent engines
- Bundled Claude Code CLI and OpenAI Codex app-server (shared/engine-capabilities.ts)
- Interface
- Desktop app with integrated terminal and Monaco editor (package.json)
- Credential storage
- API keys encrypted with Electron safeStorage, never sent to the renderer (src/main/api-key.ts)
- CI scope
- Lint and build on Linux; tests not run in CI (.github/workflows/ci.yml)
- Paid features
- Sync, backup and phone control in a separate private repository (README.md)
Read from README.md, package.json, LICENSE, src/main/index.ts, src/main/guardrails.ts, src/main/auth.ts, src/main/api-key.ts, shared/model-catalog.ts, shared/delegation.ts, shared/engine-capabilities.ts, src/main/contained-read.ts, electron-builder.yml, .github/workflows/ci.yml.
What it can do
Build software using an existing Claude or ChatGPT subscription instead of API credits
Claude or ChatGPT account credentials → AI coding agent access
Manage a real project folder with file previews
Project folder → File previews
Write code via AI agent
Natural language instructions → Source code
Run code via AI agent
Project code → Execution results
Create versioned save points of a project
Project state → Save point/version
Undo agent actions
Save point/version → Reverted project state
Approve or reject agent actions step-by-step
Proposed agent action → Approved or rejected action
Apply guardrails to agent actions
Agent action request → Constrained/limited action
Tags
Tech Stack
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.