kode · 5 MB zero-dependency Go agent runtime

The Five Whys

Why? Because every AI agent framework today ships a 200-package dependency tree and a 200 MB Python venv. kode is a 5 MB static binary with one indirect dependency.
Why? Because you shouldn't need a PhD in Python packaging just to run go test. kode installs in 5 seconds with a single go install command — no Docker, no venv, no npm, no sagas.
Why? Because locking into one LLM provider is a strategic mistake. kode is provider-agnostic — Deepseek, OpenAI, Anthropic, Ollama, vLLM, Groq, Together, Fireworks. Any OpenAI-compatible endpoint.
Why? Because agents need real tools, not abstractions on top of abstractions. kode gives you read_file, write_file, patch, search_files, shell, and browser — all gated by a unified security layer. No plugins. No SDKs. No class hierarchies.
Why? Because the agent ecosystem has split into two worlds — tools that serve MCP (Claude Code, Cursor) and tools that consume MCP. kode does both in one binary. Serve your tools to any MCP client, or connect to external MCP servers for Playwright, Fetch, databases, and more.

What you get

🔒

Sandboxed Execution

kode run --sandbox — every session spawns an isolated Docker container. No network, no host mounts beyond the working directory, zero capabilities, destroyed on exit.

🧩

Sub-Agent Delegation

Parallel OS-process sub-agents via delegate_tasks. True isolation — each sub-agent is a fresh process with its own config, tools, and timeout. Up to 8 concurrent workers.

🧠

Self-Learning Skills

Detects multi-step procedures, error recoveries, and corrections during your sessions. Saves them as reusable SKILL.md files. On by default — --no-learn to disable.

💾

Persistent Memory

Three tiers: facts, session buffer, and episodes. Merge-on-write via go-vector cosine similarity. Saves ~80% LLM calls on repeated topics.

🔗

MCP Two-Way

Server: kode mcp exposes native tools to Claude Code, Cursor, any MCP client. Client: kode connects to external MCP servers and makes their tools available to the agent.

🌐

Web UI

kode serve — browser-based agent with @ resource completion, WebSocket streaming, and a full IDE-style console. Built from Go's embed — zero npm.

🏗️

Layerable Config

Four-layer priority chain: global → project → KODE_* env vars → CLI flags. ${VAR} substitution in config files. Sensible defaults at every layer.

🔍

Native Tools

Built-in read_file, write_file, search_files, patch, shell, and browser. All gated by a unified security layer — classify operations as allow / deny / prompt per risk class.

kode vs the rest

	kode	Python agents
Dependencies	Zero. stdlib only.	200+ packages
Binary size	~5 MB static	50–200 MB (venv)
Startup	Instant	2–10 seconds
Sandbox	`--sandbox` flag	Manual Docker setup
Tool interface	One interface, one method	Class hierarchies + decorators
MCP	Bidirectional (server + client)	Client-only or adapter needed
Language	Go — native binary	Python — interpreter required

Install in 5 seconds

1 Have Go 1.24+ installed. If not: brew install go or download.
2 Run: go install github.com/BackendStack21/kode/cmd/kode@latest
3 Set your API key: export DEEPSEEK_API_KEY=sk-... or OPENAI_API_KEY
4 Run your first task: kode run "How many lines in go.mod?"

# Results instantly
$ kode run "How many lines in go.mod?"
3 lines

# With session persistence
$ kode run --session "Refactor the auth module"
$ kode continue "Now add error handling"

# Different provider, same binary
$ kode run --model gpt-4o --base-url https://api.openai.com/v1 "Explain this"

Cheatsheet

Commands

kode run <task>Single-shot task
kode run --session <task>Save conversation as session
kode continue [--id <id>] <task>Resume a saved session
kode replInteractive multi-turn REPL
kode session listList recent sessions
kode session show [id]View session transcript
kode session delete <id>Delete a session
kode session cleanup <days>Delete old sessions
kode skill listList available skills
kode skill import <uri>Import skill from URL or file
kode skill curateAudit skill quality and overlap
kode serve [--addr :8080]Start Web UI server
kode subagent --goal <s>Run a focused sub-task
kode mcp [--sandbox]Start MCP server
kode init [--global]Create config file

Key Flags

--model <name>LLM model (deepseek-v4-flash, gpt-4o, etc.)
--base-url <url>API endpoint URL
--sandboxRun in Docker sandbox
--thinking <level>Reasoning depth (enabled/disabled/low/medium/high)
--learnEnable skill learning (on by default)
--no-learnDisable skill learning
--max-iter <n>Max think→act cycles (default 90)
--system <prompt>Override system prompt
--no-colorDisable colored output

Everything else is documented

Deep-dive docs for every subsystem. Real detail, no fluff.

CLI ReferenceAll commands, subcommands, flags, examples ConfigurationConfig files, env vars, priority chain Providers & ModelsDeepseek, OpenAI, Anthropic, Ollama, vLLM, more MCP (Two-Way)Serve tools + connect to external MCP servers Memory SystemThree-tier facts, buffer, episodes Multi-Turn SessionsSave, resume, list, trim, cleanup SandboxingDocker isolation, config, security model SecurityPrompt injection defense, threat model Sub-AgentsTask decomposition, delegation protocol Self-LearningSkill detection, LLM enhancement, curation Web UIkode serve, WebSocket protocol, @ resources DevelopmentBuilding, testing, contributing

Ship it

One binary. One loop. Zero frameworks.
Go install and go.

go install github.com/BackendStack21/kode/cmd/kode@latest GitHub →