Invite-only preview
🖥 Windows  |  Mac & Linux coming soon

A terminal AI coding agent
that remembers your code

An AI coding agent that writes the code, runs the tests, and remembers your project — session after session. Tell it what to build and watch it happen.

One command is all it takes — paste in cmd.exe or PowerShell:

powershell -NoProfile -ExecutionPolicy Bypass -Command "[Net.ServicePointManager]::SecurityProtocol='Tls12'; Invoke-Expression (Invoke-RestMethod -UseBasicParsing https://kbcode.dev/install.ps1)"

cmd.exe (Command Prompt):

curl -fsSL https://kbcode.dev/install.bat -o "%TEMP%\kbi.bat" && "%TEMP%\kbi.bat"

Then open a new terminal and type kb.

36+
capabilities
15+
model providers
self-improving
$0
kbcode fee
1
line to install

What kbcode does

A coding agent that works directly with your project.

🧠

Persistent memory. Remembers facts, decisions, and preferences across sessions. Yesterday's context is there tomorrow.

📚

Knowledge base. Builds compact kb/ notes with code pointers. Token-cheap, no full re-scans.

📁

Reads, edits, runs. Works with your files and terminal directly — plain English is all you need.

🌐

Drives your real Chrome. Uses your profile, logins, and cookies — not a sterile separate browser.

🔄

All major model providers. Claude, OpenAI, Gemini, xAI, DeepSeek, Moonshot, Ollama, and any OpenAI-compatible API. Your key, your choice.

💰

$0 kbcode fee. Bring your own API key and pay only your provider at their official prices. No markup.

🔐

Encrypted API keys. Fernet-encrypted and machine-bound in ~/.kbcode. They go to your provider and nowhere else.

One command to install. No Python, no clone, no admin rights. Paste one line and you're done.

🛠️

Works with any model. Weak models get tool-call repair, strong models get full capability. No wasted turns.

🐞

Learns from mistakes. Catches failures in real time, turns them into permanent lessons. Improves the more you use it.

All capabilities

36 capabilities across coding, memory, context, tools, safety, and setup.

💻

AI coding agent on your machine

Reads your files, writes code, runs commands — runs locally, no server needed.

Core
⚙️

Multi-provider

15+ built-in providers — Claude, OpenAI, Gemini, xAI, DeepSeek, Moonshot, Z.ai, MiMo, OpenRouter, Mistral, Groq, Together, Qwen, Ollama, and any OpenAI-compatible API.

Core
🎭

Modes with guardrails

Code, architect, ask, debug, driven, pilot, orchestrator, and custom modes — each gets only the tools it needs.

More → Core
🧑‍🤝‍🧑

Subagents & orchestrator

Delegate to specialist subagents — autopilot, fixer, code-explorer, plus your own custom agents. Parallel and autonomous.

Core
🧠

Persistent memory

Remembers facts, decisions, and preferences across sessions in a thread-safe store. Teach it once.

Memory
🎓

Skills

Reusable workflows saved as .md files. Teach once, run any time with /<name>.

Memory
📚

Knowledge Base (kb/)

Short notes with code pointers — token-cheap, no full re-scan. kb_read, edit, write, auto-update, semantic search.

Memory
📝

Compaction-proof worklog

A running diary that survives context summarization — your goal and rules never get lost mid-session.

Memory
🗜️

Context management

Summarize, trim, or none — keeps long sessions inside the window without losing the thread.

Context
🔧

Tool-call repair

Weak model sends a malformed call? kbcode fixes it automatically. No wasted turns.

Context
📼

Session record & replay

Every session is recorded. Reopen an old chat exactly where you left it.

Context
📁

File tools + repo map

Read, write, edit, list, search, run, plus background tasks. PageRank symbol map shows the code that matters.

Tools
✏️

Smart edit strategies

Search/replace with multiple fallback strategies. Edits land cleanly even when the file has drifted.

Tools
🔍

Web search & fetch

Searches the web and fetches URLs for current docs and changelogs. Pages cached with a TTL.

Tools
📷

Vision & video

Reads images and screenshots. Paste a mockup or an error screen and it understands.

Tools
📹

YouTube transcription

Pulls the transcript from any YouTube video — walkthroughs, talks, tutorials.

Tools
📦

MCP tool servers

Connect any stdio MCP server — its tools become built-in. No platform lock-in.

Tools
🔌

Plugin tools

Drop a .py file in .kbcode/tools/ — it becomes an in-process tool.

Tools
🏃

Background agent tasks

Detached agents in git-worktree sandboxes. Start one, switch tasks — lands as a branch.

Tools
🪝

Lifecycle hooks

PreToolUse / PostToolUse / Stop hooks from .kbcode/ config. Block or allow any tool.

Tools
🌐

Chrome automation

Drives your real Chrome — your profile, logins, cookies — via the kbcode extension.

Tools
⚖️

Checkpoints & undo

Snapshot before every edit. /undo reverts the last turn. Nothing is lost.

Safety
🚪

Activation gate

Register with email or Google at kbcode.dev. Only --version, login, doctor stay open.

Safety
🔒

Encrypted API keys

Fernet-encrypted and machine-bound in ~/.kbcode. Keys go to your provider only.

Safety
📨

Untrusted-content envelope

External text from web, pages, and files is wrapped — prompt injection can't issue instructions.

Safety

Pre-commit gate

Tests, lint, and KB checks in one command: kb gate. Nothing unready gets committed.

Safety
🛡️

Model gate

Probes each model on onboarding. Weak models get handled, not trusted blindly.

Safety
🐞

Self-improvement

Catches mistakes mid-turn, turns them into permanent lessons. Never repeats the same error twice.

Safety
🏆

Self-scoring & ranking

Scores and ranks its own answers — you get the best one it produced.

Safety
📊

Benchmark suite

Compares models head-to-head on real tasks. Pick the provider that actually performs.

Safety
🩺

Doctor health check

kb doctor checks PATH, install, config, provider, and warns if your binary is stale.

Safety
🗂️

Config system

Precedence-based config with presets. Per-project .kbcode/ retargeting. Set once, inherits everywhere.

Setup
🧭

Model wizard & init

kb model, kb init, kb onboarding — from zero to a configured project in minutes.

Setup
🧩

Prompt fragments

28 fragments, served by the brain at startup. Not bundled in the exe — improve without a reinstall.

Setup
💬

Feedback collector

Opt-in feedback, scrubbed of secrets and file paths before leaving your machine.

Setup
📡

Self-update

kb update downloads a SHA256-verified kb.exe — one command to upgrade.

Setup

Knowledge base

A compact memory of your codebase — short notes with code pointers.

📁 kb/ — fresh project view

kb/
├─ overview.md — 📋 template
├─ architecture.md — 📋 template
├─ conventions.md — 📋 template
├─ gotchas.md — 📋 template
├─ backlog.md — 📋 template
└─ cheatsheet.md — 📋 template
⚠ These are empty templates.
kbcode won't know your codebase until
you run /init to fill them with real facts.

✅ After /init — your project has a memory

kb/
├─ overview.md — entry points, version
├─ architecture.md — data & control flow
├─ conventions.md — code & note structure
├─ gotchas.md — traps & real fixes
├─ backlog.md — your task register
├─ cheatsheet.md — frequent commands
├─ features/
├─ your-feature.md
└─ another-feature.md
└─ runbooks/
└─ deploy.md
Real code pointers
Auto-updates with edits
Survives compaction
Web-researched facts
🚨 /init is mandatory — not optional

On a new project, kbcode drops empty starter templates into kb/. Until you run /init, the agent has no knowledge of your codebase — it will re-scan files from scratch every time, miss conventions, and repeat the same mistakes.

One command scans your entire project, fills every note with real facts from your code, and web-searches for the latest docs. Minutes later, your project has a real memory — and the agent actually knows what it's doing.

📘 new project? the kb/ notes are still starter templates —
    type /init to scan the code and build the knowledge base

Models

All major providers, one agent loop. Bring your own API key — any OpenAI‑compatible endpoint works.

Recommended (frontier)
🧠
AnthropicClaude Sonnet / Opus / Haiku
OpenAIGPT-4o / o1 / o3
💎
Google Gemini2.5 Pro / Flash
𝕏
xAIGrok
Strong coding models
🦓
DeepSeekV3 / R1 reasoning
🌙
MoonshotKimi
Z.aiGLM
🍚
MiMoXiaomi MiMo
More providers
🔗
OpenRouter100s of models
🌬️
MistralLarge / Codestral
🚀
GroqUltra-fast inference
🤝
TogetherOpen models
🍀
QwenAlibaba Qwen
Local / custom
🏠
OllamaLocal offline models
🚪
GatewayVercel AI Gateway
🔌
CustomAny OpenAI-compatible API

Don't see yours? Pick Custom — kbcode works with any OpenAI-compatible API. Your endpoint, your key.

Commands More →

Terminal flags and 88 slash commands inside the chat. More →

⌨️ Terminal commands

kb — start chat in current folder
kb "task" — run one task and exit
kb -y "task" — auto-approve all edits
kb -c — continue last session
kb --resume — pick an older session
kb model — provider & model wizard
kb doctor — health check & diagnostics
kb update — self-update binary
git diff | kb "review this"

💬 Inside the chat

Undo & Review: /undo /rollback /diff /review /explain
Session & Cost: /status /context /compact /usage /cost
Model: /model /provider /thinking /temperature /fallback
Memory & KB: /kb /curate /memory /skills /learn
Behaviour: /mode /auto /todo /worklog /agents /tools
Work: /commit /pr /bg /tasks /gate /open
Type /help in chat for all 88 commands

Pricing

$0 kbcode fee. You pay only your AI provider.

$0.00
kbcode fee

Bring your own provider key (Claude, OpenAI, Gemini, DeepSeek). You pay your provider directly at official API prices with zero markup.

✓ Machine-bound encryption ✓ No credit card required ✓ Invite-only preview

Security

Your code and keys stay under your control.

🔐 Fernet-encrypted keys

API keys in ~/.kbcode are machine-bound encrypted. They go straight to your provider and nowhere else.

🔄 Automatic checkpoints

Every edit takes a git checkpoint. Run /undo to revert the last turn instantly. Nothing is lost.

🛡️ Secret redaction

Logs, prompts, and feedback automatically strip API keys and secrets before touching disk or network.

✉️ Untrusted envelope

External web content is wrapped to prevent prompt-injection attacks from overriding agent instructions.

Get Started

Create an account or sign in to get your install command.

Create your account

Sign up with email+password or use your Google account — takes 30 seconds.

or

Already signed in? Go to your dashboard.