# Kite

> Kite 1.0.6 is an open source terminal coding agent for Python 3.11 or newer. It runs a model against a local repository through 19 coding tools and 27 tools in total, records the result of each check, and blocks submission until there is verification evidence. MIT licensed, 16 model providers, on macOS, Linux, Windows, WSL.

This file is the machine-readable summary of https://kitedev.vercel.app. It is generated from the Kite repository at commit a96f8bf (2026-10-05), and every version, count, name, command and limit below is read out of that repository rather than written by hand, so it cannot drift from the tool it describes.

## What Kite is

Kite is a terminal coding agent. It runs a language model against a local repository, in a session you can read and steer, using the model you already pay for. It is not a hosted service and it runs no telemetry. The repository describes itself as: "Slim hybrid coding-agent harness (mini-swe-agent loop + tau-style tools/events/providers/memory)"

- Runs on: macOS, Linux, Windows, and WSL.
- Written in Python, and needs version 3.11 or newer to run.
- Licensed MIT, by Uzeb Khan.
- Current version: 1.0.6, released 2026-10-05.

## Key facts

- Name: Kite
- Version: 1.0.6
- License: MIT
- Author: Uzeb Khan (https://uzebkhan.vercel.app)
- Source: https://github.com/KhanUzeb/kite
- Released: 2026-10-05
- Requires: Python >=3.11
- Platforms: macOS, Linux, Windows, WSL
- Model providers: 16
- Coding tools: 19
- Total tools: 27
- Modes: plan, build
- Bundled skills: 7
- Slash commands: 4
- Effect classes that always need a decision: 5
- Crews: 3 workers, cap 12, depth 1
- Blocked shell patterns: 46
- Default steps and cost per task: 40, $5.00
- Repository lines of Python: 55,786
- Repository commit read for these facts: a96f8bf (2026-10-05)

## Install

Kite installs from a script in a public repository. It is not on a package registry.

macOS, Linux, and WSL:

```sh
curl -fsSL https://raw.githubusercontent.com/KhanUzeb/kite/main/scripts/download.sh | bash
```

Windows PowerShell:

```powershell
irm https://raw.githubusercontent.com/KhanUzeb/kite/main/scripts/download.ps1 | iex
```

Windows, if the PowerShell execution policy blocks that:

```powershell
powershell -NoProfile -ExecutionPolicy Bypass -Command "irm https://raw.githubusercontent.com/KhanUzeb/kite/main/scripts/download.ps1 | iex"
```

Then run `kite setup` to choose a provider and a default model, and run `kite` inside a repository.

## What the tool does

- Edits a repository through 19 coding tools: read, write, edit, bash, grep, glob, ls, set_cwd, skill, todo_write, todo_read, task, webfetch, websearch, webcrawl, subagent, memory, submit, question. With the optional Context7 documentation lookup (2 tools) and GitHub tools (6), 27 tools are available.
- Derives the side effects of every tool call before running it, and asks for a decision on 5 effect classes that always require one: destructive, network, durable_memory, package_or_skill_install, nested_agent. With no approver available, those calls are refused rather than run.
- Records the result of each check and blocks submission until there is evidence. The gate is on by default (true). A change is one of idle, changed_unverified, failed, verified, or partial, and only verified or partial lets a submit through.
- Runs in 2 modes. plan is read-only and blocks write, edit, bash at the dispatch layer; build allows them. plan permits 21 read, search and fetch tools.
- Delegates to crews of at most 12 coder, context, reviewer, scout, shell subagents, 3 running at a time, with a depth ceiling of 1 and a ceiling of 64 spawns per task. A worker cannot spawn further agents.
- Refuses 46 shell patterns unconditionally, including `rm -rf /`, disk writes, fork bombs, and piping a download into a shell, and grades the rest into 3 consequence tiers: routine, serious, critical.
- Ships 7 bundled skills (/commit, /debug, /init, /orchestrate, /pr, /research, /review) and 4 slash commands (/explain, /fix, /pr, /unslop), and can load more from a workspace, npm, GitHub, or a local folder.
- Keeps sessions as JSONL, resumes the last one, and discovers project context from the working directory.
- Offers 6 terminal themes.

## Limits and defaults

The shipped values in the repository's default configuration. Each one is overridable in the config, and a run that hits a ceiling reports which one rather than stopping quietly.

- Steps per task: 40
- Cost per task: $5.00
- Steps per interactive session: 80
- Cost per interactive session: $10.00
- Shell timeout: 120s
- Provider call retries: 4
- Output kept: 16,384 characters of `bash` output, 5,000 characters of tool observation
- Loop hard threshold: a task that spends 7 loops without progress is stopped as stuck
- Verification before submit: true
- Execution mode: host, with the working directory sandboxed unless changed (sandbox_to_cwd = true)

## Model providers

Kite ships 16 provider entries and hardcodes no model names, so a new provider release works without a Kite update. Auth metadata comes from your user config or the environment, never from the repository you are working in.

- Subscription: ChatGPT (subscription), Claude (subscription), Grok (subscription), Antigravity (subscription), OpenCode Go.
- API key: OpenAI, Anthropic, xAI, OpenRouter, Hugging Face, Google Gemini, Groq, OpenCode Zen, NVIDIA NIM.
- Local: Ollama (local).
- Custom: Custom OpenAI-compatible.

Caveats that apply to specific providers:

- **ChatGPT (subscription)**: OAuth — uses your ChatGPT/Codex subscription, not API credits
- **Claude (subscription)**: Claude Code can be linked for status; Kite model calls need ANTHROPIC_API_KEY
- **Grok (subscription)**: OAuth — uses your SuperGrok plan, not API credits
- **Antigravity (subscription)**: OAuth — chat runs on your Google subscription via the signed-in agy CLI; GEMINI_API_KEY adds tool-driven agent runs
- **OpenCode Go**: Flat monthly subscription — not per-token API billing

## What Kite does not do

- MCP stdio servers were removed. Legacy `[[mcp]]` keys in runtime TOML are ignored rather than silently half-working.
- Context7 is the only bundled documentation integration.
- There is no model allowlist, so Kite will attempt a model name it has never seen.
- Subagent depth is 1, so a crew is a bounded fan-out rather than a recursive tree.
- Secret redaction is pattern-based, not exhaustive.
- Tool calls run on your own machine, not on a remote host. Execution mode is host and the boundary is the working directory.

## Verify these claims yourself

Nothing here is asserted on the authority of this site. Each value was read out of the Kite repository, and each one can be read again.

- Repository: https://github.com/KhanUzeb/kite
- Commit these facts were read from: a96f8bf
- Date of that commit: 2026-10-05
- Release date of 1.0.6: 2026-10-05
- Tests: https://github.com/KhanUzeb/kite/tree/main/tests, runnable with `uv run pytest` and no model key
- Lines of Python under `src/`: about 55,786
- Release notes: https://github.com/KhanUzeb/kite/blob/main/CHANGELOG.md
- License text: https://github.com/KhanUzeb/kite/blob/main/LICENSE

## Pages

- [Product overview](https://kitedev.vercel.app/): Current capabilities, each linked to the file in the repository that implements it.
- [Documentation](https://kitedev.vercel.app/docs): Install commands, model providers, plan and build modes, approval, tools, crews, verification, skills, and limits.
- [Changelog](https://kitedev.vercel.app/changelog): What shipped in the current release.

## Source

- [Repository](https://github.com/KhanUzeb/kite): The implementation, and the project documentation, behind every number here.
- [CHANGELOG](https://github.com/KhanUzeb/kite/blob/main/CHANGELOG.md): Every release, including 1.0.6.
- [LICENSE](https://github.com/KhanUzeb/kite/blob/main/LICENSE): MIT license text, naming Uzeb Khan.
- [Test suite](https://github.com/KhanUzeb/kite/tree/main/tests): The tests behind every claim above, runnable without a model key.

## Optional

- [Author and maintainer, Uzeb Khan](https://uzebkhan.vercel.app): The person named in the LICENSE file.