Kon is a minimal coding agent with a tiny harness: about 215 tokens for the system prompt and around 600 tokens for tool definitions – so under 1k tokens before conversation context.

At the time of writing this README (25 Feb 2026), this repo has 112 files and is easy to understand in a weekend. Here’s a rough file-count comparison against a couple of popular OSS coding agents:

Others are of course more mature, support more models, include broader test coverage, and cover more surfaces. But if you want a truly minimal coding agent with batteries included – something you can understand, fork, and extend quickly – Kon might be interesting.

$ fd . | cut -d/ -f1 | sort | uniq -c | sort -rn
4107 opencode
 740 pi-mono
 108 kon

Kon is inspired from Bleach, a artificial soul

Python 3.12+ and uv.

uv tool install kon-coding-agent

This installs kon globally as a CLI tool.

git clone https://github.com/kuutsav/kon
cd kon
uv tool install .

kon

CLI options:

usage: kon [-h] [--model MODEL]
           [--provider {github-copilot,openai,openai-codex,openai-responses,zhipu}]
           [--api-key API_KEY] [--base-url BASE_URL] [--continue]
           [--resume RESUME_SESSION]

Kon TUI

options:
  -h, --help            show this help message and exit
  --model, -m MODEL     Model to use
  --provider, -p {github-copilot,openai,openai-codex,openai-responses,zhipu}
                        Provider to use
  --api-key, -k API_KEY
                        API key
  --base-url, -u BASE_URL
                        Base URL for API
  --continue, -c        Resume the most recent session
  --resume, -r RESUME_SESSION
                        Resume a specific session by ID (full or unique
                        prefix)

• fd – required for fast file discovery; Kon auto-downloads it only if it's missing.
• ripgrep (rg) – required for fast content search; Kon auto-downloads it only if it's missing.
• eza (optional) – supports .gitignore-aware listings and usually emits fewer tokens than ls.

• GitHub Copilot OAuth: run /login and choose GitHub Copilot.
• OpenAI OAuth (Codex): run /login and choose OpenAI. Kon supports callback flow plus manual paste fallback.
• OpenAI-compatible providers (for example ZhiPu): set an API key via environment variable (OPENAI_API_KEY or ZAI_API_KEY).

Type / at the start of input to see available commands.

Type @ + query to fuzzy-search files/folders in the current project and insert paths into your prompt.

Press Tab in the input box to complete paths (~, ./, ../, absolute paths, quoted paths, etc.).

If the agent is currently running, you can still submit more prompts. Kon queues them and runs them in order once the current task finishes (up to 5 queued prompts).

Sessions are append-only JSONL files under ~/.kon/sessions/.

• /resume to reopen past sessions
• /session for message/token stats
• /export for standalone HTML transcripts
• --continue / -c to continue the most recent session from CLI

Kon loads project guidelines from AGENTS.md (or CLAUDE.md) files into the system prompt:

1. Global: ~/.kon/AGENTS.md
2. Ancestor directories from git root (or home) down to current working directory

Skills are reusable instruction packs loaded from:

• Project: .kon/skills/
• Global: ~/.kon/skills/

Each skill has a SKILL.md file with front matter:

---
name: my-skill
description: Brief description of what this skill does
---

# My Skill

Detailed instructions for the agent...

For skills with scripts, see Agent Skills Documentation.

Some features you might expect in other coding agents are not part of Kon's design philosophy:

• MCP servers – Use skills instead; they're simpler and give you full control
• Sandbox environments – Kon runs directly on your machine for simplicity; use Docker or VMs if you need isolation
• Checkpoint restores – Not currently supported; may be added in the future

LLM Provider
    │
    │ StreamPart (TextPart, ThinkPart, ToolCallStart, ToolCallDelta, ...)
    ▼
Single Turn (turn.py)
    │
    │ StreamEvent (ThinkingStart/Delta/End, TextStart/Delta/End, ToolStart/End, ToolResult, ...)
    ▼
Agentic Loop (loop.py)
    │
    │ Event (AgentStart, TurnStart, TurnEnd, AgentEnd + all StreamEvents)
    ▼
UI (app.py)

Kon works well with local models exposed through an OpenAI-compatible /v1 API.

To run a local model using llama-server:

./llama-server -m <models-dir>/GLM-4.7-Flash-GGUF/GLM-4.7-Flash-Q4_K_M.gguf \
    -n 8192 \
    -c 64000

# Then use Kon with:
kon --model zai-org/glm-4.7-flash \
    --provider openai \
    --base-url http://localhost:8080/v1 \
    --api-key ""

GLM-4.7-Flash-Q4 ran at 80-90 tps on my i7-14700F × 28, 64GB RAM, 24GB VRAM (RTX 3090)

Config lives at ~/.kon/config.toml (auto-created on first run).

Most important knobs:

• llm.default_provider
• llm.default_model
• llm.default_thinking_level
• llm.system_prompt (you can fully override Kon’s system prompt here)
• llm.tool_call_idle_timeout_seconds (fallback timeout for stalled tool-call streaming)
• compaction.on_overflow, compaction.buffer_tokens
• agent.max_turns, agent.default_context_window

You can also theme the UI via [ui.colors] values.

Example:

[llm]
default_provider = "openai-codex"
default_model = "gpt-5.3-codex"
default_thinking_level = "high"
tool_call_idle_timeout_seconds = 60
system_prompt = """Your custom system prompt here"""

[compaction]
on_overflow = "continue"
buffer_tokens = 20000

For hacking on Kon locally:

uv sync
uv run kon
uv run ruff format .
uv run pytest

• Kon takes significant inspiration from pi-mono coding-agent, especially in terms of the overall philosophy and UI design.
  • Why not just use pi? Pi is no longer a small project, and I want to be in complete control of my coding agent.
  • I mostly agree with Mario (author of pi), but I have different beliefs on some matters - for example, subagents (especially useful for context gathering in larger repos when paired with semantic search tools).
  • Over time, I also want to give more preference to local LLMs I can run. glm-4.7-flash and qwen-3-coder-next look promising, so I may make decisions that do not necessarily optimize for SOTA paid models.
• Kon also borrows ideas from Amp, Claude Code, and other coding agents.

MIT