Implant
Implant lets AI coding assistants read and edit your code with the same intelligence
your editor already has — every language server, every extension command, plus
ast-grep for cross-language structural search and rewrites.
Under the hood, agents send short JavaScript snippets that run inside the VS Code
extension host with the full vscode.* API in scope. Nothing runs without the user
approving the exact snippet.
Explore the code with the tools the editor already has: Find References,
Go-to-Definition, hover types, workspace diagnostics, git blame, symbol outlines,
and any command a third-party extension registers.
Modify the code through the same language server that powers the editor:
structural rename, quick-fixes and refactors, organize imports, safe file
create/rename/delete via WorkspaceEdit, formatter output, source actions. Edits
go through the language server — safer than sed/awk on files because they
respect scope, imports, and semantics.
Structural search and rewrite with ast-grep: find and change every occurrence
of a code shape across the workspace at once, in any language ast-grep supports,
without the false hits regex gives you on comments and string literals.
Quick start
- Install the extension.
- Open your project. Implant will offer to install itself into the workspace —
this drops portable rule files that teach agents when to reach for Implant.
- Your agent picks up the tool automatically — MCP-native clients (Copilot Chat
and others on VS Code ≥ 1.101) discover it via the registry, and CLI agents
pick it up from the workspace instructions installed in step 2.
- When an agent invokes the tool, a webview opens beside the editor showing the
script with syntax highlighting. Click Run, Run + trust session, or
Deny.
You can also ask Implant directly in Copilot Chat via @implant — see below.
@implant in Copilot Chat
Type @implant … in Copilot Chat with a natural-language request. Implant asks
the language model to translate your ask into a vscode.* snippet, shows you
the snippet, runs it through the same approval flow as any other invocation,
and streams the result back into the chat.
Examples of things that aren't one-click in the UI:
@implant list every TODO / FIXME / HACK comment across the workspace, grouped by tag, with clickable file:line links.@implant which exported symbols in src/ have zero references? Those are dead-code candidates.@implant sort my open tabs by diagnostic count so I know which file to fix first.@implant give me a bug-report snapshot: VS Code version, active user extensions, open tabs, dirty files, and per-language diagnostic count. Format it as GitHub-ready Markdown.@implant list the top 20 functions in this workspace by cross-file reference count.@implant preview what a project-wide rename of \foo` → `bar` would touch — file count and line count per file, no changes applied.`
Requires GitHub Copilot for the language model. Read-only queries and refactor
dry-runs are the sweet spot; anything that modifies files still opens the
approval webview so you can eyeball the exact snippet before it runs.
Safety notes
- Snippets run in the VS Code extension host with the same privileges as the
extension itself. There is no sandbox — that is the entire point.
- The approval webview exists so you can eyeball each snippet before it runs.
Turn confirmation off only when the agent is already in a tight review loop.
- The HTTP server only binds to
127.0.0.1 and requires a per-session bearer
token — the token lives in .vscode/implant/session.yml (mode 0600, gitignored,
regenerated on every activation). Requests without a valid token get 401 and
raise a warning notification so you notice anything poking at the port.
- Even with the token gate, don't run this on a shared machine: any process
running as your user can read the session file.
- The
implant.requireConfirmation workspace setting toggles the approval
webview. Leaving it on is the default for a reason; the per-session "Trust"
button on the approval webview is the escape hatch when you're actively
supervising a run.
