Kiro — AI Coding Agent
Kiro is a real AI agent that lives inside VS Code. It reads and edits your files,
runs commands in the integrated terminal, verifies the result with evidence, and asks
before risky actions.
Independent project. Not affiliated with, endorsed by, or connected to
Amazon Web Services or its "Kiro" product.
Why Kiro?
- Bring your own key — Google Gemini, OpenRouter, Groq, NVIDIA, Mistral, xAI,
Qwen, OpenAI, Anthropic, or a local model through Ollama / LM Studio.
- Real work, not autocomplete — the agent writes files, edits code, runs shell
commands, searches the workspace, and reads the output before it reports success.
- Evidence-based finishing — a task is only marked successful after Kiro checks
the file or command output. No "done" without proof.
- Approval modes —
balanced (default) asks before destructive or external
actions, strict asks for every step, yolo runs autonomously.
- Workspace aware — attaches the active file and selection to your message, and
supports per-folder rules through
KIRO.md / AGENTS.md.
- Skills, MCP, and checkpoints — extend the agent with GitHub skills and MCP
servers, and roll back file edits from the built-in snapshots.
Quick start
- Install the extension and open the Kiro icon in the activity bar
(or press
Ctrl+Alt+K).
- Pick a provider and model in the chat header.
- Click the key icon and paste your API key — it is stored locally on your machine.
- Describe the task and press Enter.
Providers
| Provider |
Notes |
| Google Gemini |
Recommended default, generous free tier |
| OpenRouter |
Use models ending in :free on a free account |
| Groq / NVIDIA / Mistral / xAI / Qwen |
Bring your own key |
| Ollama / LM Studio |
Fully local, no key |
| OpenAI / Anthropic / custom |
Any OpenAI-compatible endpoint |
Commands
| Command |
Description |
Kiro: Open Chat |
Focus the Kiro chat view |
Kiro: New Chat |
Clear the conversation |
Kiro: Stop Task |
Stop the running task |
Kiro: Explain Selection |
Explain the current selection |
Kiro: Fix or Improve Selection |
Fix or improve the current selection |
Kiro: Choose Provider and Model |
Pick provider and model |
Kiro: Set Provider and Key |
Set the API key |
Kiro: Connection Status |
Show engine status |
The chat can also be opened from outside VS Code:
vscode://kiro-pc.kiro-agent/chat?task=<url-encoded task>
Settings
| Setting |
Default |
Description |
kiro.provider |
saved config |
Default provider |
kiro.model |
saved config |
Default model |
kiro.mode |
balanced |
Approval mode: yolo, balanced, strict |
kiro.maxSteps |
40 |
Maximum steps per task |
kiro.verify |
false |
Evidence verification before finishing |
kiro.sendSelection |
true |
Attach the active file/selection |
kiro.openOnStartup |
false |
Open the chat when VS Code starts |
kiro.pythonPath |
python |
Python used when the bundled bridge is unavailable |
kiro.agentCommand |
empty |
Custom engine command |
How it works
The extension talks to the Kiro engine over a local JSONL stdio bridge. The bridge is
bundled as bin/KiroBridge.exe (self-contained, no Python required on Windows). If the
binary is not available, the extension falls back to python -m kiro.bridge using the
bundled Python package. No network port is opened, and the engine never leaves your
machine except for model API calls.
Privacy and security
- API keys are stored locally in your user profile and are masked in every log line.
- Nothing is sent anywhere except to the model provider you choose.
- Destructive actions require approval in
balanced/strict modes.
- Screen text is treated as data, not instructions, to reduce prompt-injection risk.
Requirements
- VS Code 1.90 or newer.
- A model provider key, or a local model via Ollama / LM Studio.
- Windows for the bundled bridge binary; other platforms can use Python 3.10+.
Support
Open an issue at https://github.com/Kiro-404/kiro-pc/issues. Never include API keys
in a bug report.
License
MIT
| |
