Cerebretron Code

Cerebretron's AI coding agent, inside VS Code. It reads your workspace, edits files, runs commands, searches your codebase, and taps Cerebretron's cloud tools (engines, FEA, math, web) — all from a streaming chat in the sidebar or a full editor tab.

Highlights

Agentic editing — reads files to answer, makes targeted search/replace edits (not whole-file rewrites), runs commands, searches the whole codebase, and shows every change as a diff you approve.
Permission modes — Ask first (approve each step), Auto-edit (apply straight to files, review the diff), Plan (investigate → propose a plan before touching code), YOLO (run unattended).
Effort — Standard (fast), Think (deeper reasoning, streamed live), Max (a multi-model relay: write → review → audit for the highest quality).
Live feedback — a running activity indicator with a live token counter, a Writing → Reviewing → Auditing progress stepper for Max, and a "Thought for Ns" panel when the model reasons.
Deep context — CEREBRETRON.md project rules, @file / @git / @problems / @folder / @symbol mentions, language-server navigation, and multi-turn memory.
Checkpoints + Rewind — every edit is snapshotted; revert any step while keeping the conversation.
Sessions — multiple conversations per workspace.
MCP client — connect your own Model Context Protocol servers; their tools become agent tools.
Editor-power tools — inline edit (Cmd/Ctrl+K Cmd/Ctrl+I), "Fix with Cerebretron" quick-fixes, ghost-text completions, slash commands.
Rich chat — syntax-highlighted code (Copy / Apply), inline diagrams & images, full-width editor tab.

Install

From the VSIX: Extensions panel → ⋯ → Install from VSIX… → pick cerebretron-code-<version>.vsix. Or:

code --install-extension cerebretron-code-<version>.vsix

Then reload VS Code and open Cerebretron Code from the Activity Bar.

Click Sign In (or run Cerebretron: Sign In) → Sign in with browser — authorize in your browser, no key to copy. (Pasting a gns_ API key from cerebretron.ai still works; it's stored in VS Code SecretStorage, encrypted, never in settings.)

Using it

Chat: Enter sends · Shift+Enter for a newline · Send → Stop cancels.
Open in an editor tab: the ⤢ header button, the Cerebretron status-bar item, or Cerebretron: Open Chat in Editor.
Pick Mode & Effort: the Mode and Effort buttons in the chat bar.
Reference context: type @ for a file/symbol/git/problems picker.
Slash commands: type / — /explain /fix /tests /doc /review, /plan, /init (generate CEREBRETRON.md), /model, /usage, /context, /compact, /mcp, /clear, /help, plus your own from .cerebretron/commands/*.md.
Inline edit: select code → Cmd/Ctrl+K Cmd/Ctrl+I → describe the change → review the diff.
Attach an image: paste a screenshot or click ＋, then ask about it.

Effort levels

Effort	What it does
Standard	Fast, single-model responses.
Think	The model reasons step-by-step before answering; its thinking streams live.
Max	A multi-model relay collaborates (write → review → audit) for the best result. Available on the Code, Reason, and Innovate tiers. Slower; opt-in.

Commands & keybindings

Cerebretron: Sign In / Sign Out
Cerebretron: New Chat
Cerebretron: Open Chat in Editor
Cerebretron: Select / Cycle Permission Mode (Cmd/Ctrl+K Cmd/Ctrl+M)
Cerebretron: Rewind Checkpoint…
Cerebretron: Sessions / History…
Cerebretron: Manage MCP Servers…
Cerebretron: Edit Selection (inline) (Cmd/Ctrl+K Cmd/Ctrl+I)
Cerebretron: Explain Selection, Toggle Inline Completions

Settings

Setting	Default	Description
`cerebretron.baseUrl`	`https://www.cerebretron.ai`	API base URL (https only).
`cerebretron.model`	`genesis-code`	Default model tier.
`cerebretron.completions.enabled`	`true`	Inline ghost-text completions.
`cerebretron.completions.model`	`genesis-flash`	Model for completions.
`cerebretron.contextWindow`	`128000`	Token window for the context % readout.
`cerebretron.mcpServers`	`{}`	Your MCP servers (http / stdio).

How the agent works

Each turn runs a tool loop in the extension host: the model emits tool calls, the extension executes them, feeds the results back, and loops until it answers.

Read-only tools (read_file, list_dir, codebase_search, LSP nav) run automatically and in parallel.
Mutating tools (edit_file, write_file, run_shell, run_in_terminal) are gated — you review a diff or the command first (subject to the permission mode).
Cloud tools (engines, CAD→FEA, math, web search, diagrams, images, knowledge) run server-side via the Cerebretron API.

Privacy & security

API keys live in VS Code SecretStorage (encrypted), never in settings.
File access is confined to your workspace (symlink-escape-proof).
stdio MCP servers run only in a trusted workspace; http MCP is SSRF/DNS-rebind hardened.
No telemetry is sent unless you explicitly opt in.

See CHANGELOG.md for version history.