Agentic tool loop
Read, list, glob, grep, run commands, and edit — with independent read-only lookups running in parallel and diagnostics fed back after every edit.
Agentic coding, on any model
The AI coding agent for VS Code that runs on any model you want.
Luna Code is an agentic assistant powered by OpenRouter. Point it at any model with just an API key and a model id — no other setup. Built for fast codebase navigation and cheap, cache-optimized long sessions.
ext install derekshreds.lunacode
- Any OpenRouter model
- Private by default
- MIT licensed
Luna Code
z-ai/glm-5.2
Thinking
Ask Luna Code to build, fix, or explain…
1-linesetup — key + model
300+models via OpenRouter
3 modesStandard · Auto · Plan
0 dataretention option
Everything in one panel
An agent that actually knows your codebase
Luna Code reads, greps, globs, runs builds and tests, checks language-server diagnostics, and edits code — looping until the task is done.
Explore sub-agent
Delegate open-ended research to a disposable sub-agent with its own context. Only the digest returns — keeping the main thread small and cache-friendly.
Usage & cost analytics
A live meter shows session cost, last-turn tokens, and cache-hit rate — plus 30/60/90-day charts and a per-model spend breakdown.
Cache-hit optimized
A byte-stable prompt prefix and rolling cache_control breakpoints
maximize provider prompt caching, so long sessions stay cheap and fast.
Turn checkpoints
Every agent turn snapshots the files it changed. One click on the ↩ revert chip restores the last turn — with a stack of 10.
Project memory
A LUNA.md at your workspace root loads into every session. The agent
records durable conventions and gotchas there — nested files supported.
MCP servers
Connect stdio Model Context Protocol servers; their tools appear to the agent as
mcp__server__tool, with approval gating in Standard mode.
Session history
Every conversation is saved per-workspace. Browse, reload, export as Markdown, or delete prior sessions from the header.
Worktree sandbox
Run the agent in a separate git worktree, review the diff, then merge or discard — your working tree stays untouched until you say so.
Secure & private
Your API key lives in VS Code's encrypted SecretStorage. Requests route only to providers that don't store or train on your prompts.
You stay in control
Three modes for how much rope the agent gets
Switch per session — from careful review to fully autonomous.
Approve every step
The agent proposes each file edit and shell command; you approve before it runs. Inline side-by-side diffs make review fast.
Fully autonomous
Runs edits and commands without prompting so it can finish long tasks — with always-deny commands hard-blocked and a session budget guardrail.
Read-only planning
Investigates the code and proposes a concrete plan, but never edits files or runs mutating commands. Mutating tools are hidden entirely.
Up and running in a minute
Setup is one API key and a model
No accounts, no config files, no per-provider wiring.
-
1
Install from the Marketplace
Add Luna Code to VS Code and open it from the Activity Bar — or pop it into an editor tab or the Secondary Side Bar.
-
2
Paste your OpenRouter key
Run Set OpenRouter API Key and paste your
sk-or-v1-…key. It's stored in encrypted SecretStorage. -
3
Pick a model & go
Choose any model from the picker, type a request, and hit Enter. Luna Code handles the rest.
Install for VS Code
Requires an OpenRouter API key.
Private & economical
Cheap long sessions, without leaking your code
Luna Code is engineered for the two things that make agentic coding painful: runaway cost and data exposure.
- Private by default. Every request sends
data_collection: "deny", so traffic only reaches providers that don't store or train on your prompts. An optional stricter Zero-Data-Retention mode is available. - Price-aware context. Cache-aware compaction keeps history append-only until a budget sized from the model's window and price is hit — then summarizes the oldest turns.
- Secure key storage. Your key lives in VS Code's encrypted SecretStorage, never in settings or files.
- Resilient streaming. Automatic retry with backoff on 429/5xx, a hung-stream watchdog, and optional fallback models.
Bring your own model. Keep your code yours.
Install Luna Code and start an agentic session in under a minute.