Lacis gives VS Code a context tree — your identity, rules and
skills, auto-injected only when relevant. Then branch your chat like
git: every exploration keeps its own clean context, and they never
bleed into each other.
The problem
Every new chat, the AI forgets who you are and how you work — so you re-paste the same context again and again. And the longer a thread runs, the dumber it gets, as unrelated detours pile into one ever-growing history.
The heart of Lacis
Define a small tree of typed branches once. On every message, a lightweight router reads each branch's “load when…” signal and pulls in only what your request actually needs — on any branch of any conversation. A banner shows exactly what was loaded, so it's never a black box.
What the model sees
git branch, but for your AI conversations. Every message is a node;
fork from any of them, and each branch keeps its own clean history.
[root]
│
┌───────────┼───────────┐
▼ ▼ ▼
[REST] [GraphQL] [gRPC]
│
▼
[JWT auth] → [refresh tokens]
Send a message from refresh tokens and the model receives exactly
root → REST → JWT auth → refresh tokens. GraphQL and gRPC never enter
the context. You can compare the branches — the model can't confuse them.
Not just chat
Lacis isn't a web page beside your work. It runs in your editor, on your repo: the agent reads, creates and edits files, and runs commands.
Bring your model
Already have Copilot (Free, Individual or Business)? Lacis uses it through the official VS Code API. No API key, no setup — install and go.
Point Lacis at OpenAI, Anthropic, OpenRouter, Groq, Mistral, or a local Ollama / LM Studio endpoint. First-class on Cursor, Windsurf and VSCodium, where Copilot isn't available.
Also in the box
Ask about the current selection, file or diagnostics. Insert a reply at the cursor. Continue a conversation from any node.
See exactly what a turn will cost — tokens and money — before you press send.
Attach an image to any message and use vision models on it.
Send a read-only conversation or context tree. Zero backend: the tree travels in the URL itself.
Tree view, focus mode, tags, bookmarks and search — navigate hundred-node conversations fast.
No account, no mandatory cloud. Conversations stay on your machine; keys stay in your editor’s secret store.
FAQ
Lacis is a VS Code extension for branching AI conversations. Every reply can fork into a new branch, and the model only ever sees the current path — never the sibling branches. On top of that, a routed context tree auto-injects your identity, rules and skills into every reply, only when relevant. It works with GitHub Copilot or any OpenAI-compatible API key, and also runs in Cursor, Windsurf and VSCodium.
ChatGPT branches a web page. Lacis branches inside your editor, on your files — with your Copilot or your own key, an approval-gated repo agent, and true context isolation per branch: each branch sends only its own ancestor chain to the model.
Editing keeps everything in the same context window, which keeps growing and drifting from the goal. A branch sees only its own path — clean context, sharper answers — and both versions coexist so you can compare them side by side.
Lacis runs on your Copilot rather than replacing it. It adds the conversation tree, per-branch context isolation, the auto-routed context tree, and multi-conversation management that a linear chat panel doesn’t have.
Not if you have GitHub Copilot (Free, Individual or Business) — Lacis uses it through the official VS Code API with zero configuration. Otherwise, bring any OpenAI-compatible key: OpenAI, Anthropic, OpenRouter, Groq, Mistral, or a local Ollama / LM Studio endpoint.
The extension is free to install. You bring your own model — your Copilot subscription or your API key — so you pay your provider directly, and Lacis shows you the token cost before every send.
Everything is local by default: no account, no mandatory cloud. Conversations are stored on your machine and API keys live in your editor’s secret store. Optional telemetry is anonymous and structure-only — never message content — and is documented in full on the transparency page.
Something else? Ask us directly — every message is read.