insrc for VS Code

Grounded code-knowledge and the tracked insrc workflow, wired into your AI assistant — the VS Code counterpart to the insrc JetBrains plugin.

insrc runs a local daemon that indexes your repositories into a graph + vector store and exposes a citation-grounded analyze/workflow toolset over a local Unix socket. This extension is a thin orchestrator: it installs and manages that daemon, wires insrc into your AI host's MCP config, and registers your workspace — all consent-gated, all reversible. It opens no cloud path and stores no secrets; every capability reaches the daemon over the local socket.

Requirements

VS Code and an AI host (Claude Code, Cursor, …) that insrc wires its MCP tools into.
Node — used by the daemon; the extension provisions a private runtime automatically if the system Node is missing or unsuitable.
Ollama — required for local embeddings (the default embedder is qwen3-embedding:0.6b; a no-Ollama ONNX alternative, nomic-embed-text-v1.5, is also available). Ollama is optional for the core model: run it locally on Ollama (qwen3.6:27b) or through your existing claude / codex CLI instead.

Full setup → Getting Started.

First run

On first activation the extension runs one coalesced onboarding flow, prompting (each step independently, and only when needed) to:

Install the insrc daemon into ~/.insrc/.
Register this workspace with the daemon (repo.add) so it gets indexed.
Wire your AI host (Claude Code, Cursor, …) so insrc's MCP tools are available to your assistant.

Skipped a prompt? Every action is also a durable command in the Command Palette:

insrc: Install daemon / Start / Stop / Restart / Update daemon
insrc: Wire AI hosts
insrc: Register workspace

Uninstalling the extension runs a vscode:uninstall hook that removes the insrc entries from every host config it wrote.

Configure insrc

The daemon's configuration is editable directly from native VS Code Settings (search insrc) — every option is a typed, machine-scoped setting, so it never syncs off this machine:

Global options (insrc.*) — log level, Ollama host, permission mode, model tiers, plan depth, embeddings, and more, each with the daemon's current value.
Per-role model tiers (insrc.models.tasks.*) — override which tier (cheap / mid / core) serves each insrc reasoning role.

The Settings view stays truthful to the daemon: an edit the daemon rejects is reverted with its reason, and external changes (from the CLI, JetBrains, or another window) are reconciled on demand via insrc: Refresh insrc settings in the Command Palette. All config access is over the local daemon socket — no cloud path.

Pick a tier's model from a dropdown — insrc: Set model tier in the Command Palette walks you through a tier → provider → model picker whose model list is exactly what the daemon offers for that tier's provider (Ollama's installed models, or the curated Claude / Codex catalog) — no typing raw model ids. If a provider can't be listed it shows no models available with a Refresh; switching the provider re-filters the list.

Keeps the backing daemon current — on activation the extension checks (only when the daemon is already reachable) whether the installed daemon has fallen behind its upstream, and if so offers an Update / Dismiss notification that updates the daemon in place on your OK. When you update the extension itself, it brings the daemon current automatically. Failures surface once with the reason; nothing is retried or rolled back — the daemon owns its own state. After a successful update it nudges you to Reload Window so the insrc MCP connection reconnects to the freshly-updated daemon (a session run outside the editor needs a manual claude / codex restart instead).

Status & panels

The status-bar insrc item opens a two-item menu (also reachable from the Command Palette):

Open Detailed Status — a tabbed panel over the running daemon:
- Daemon — the daemon's current state (uptime, repos, queue depth).
- Workflows — the workflow-artifact chain per work item.
- Debug — the attached MCP clients, a live tail of the daemon log, and a consent-gated cleanup of stray insrc processes (the managed daemon is never touched).
Open Repo Configuration — pick any repo the daemon has registered (independent of which folders are open) and edit that repo's per-repo model-tier overrides through form fields — no hand-editing JSON. Every write is consent-gated and reflects back only what the daemon actually stored.

All panels are read/edit views over the local daemon socket — no background polling (only the Debug log tails while its tab is open), no cloud path.

Building & packaging (maintainers)

cd vscode-plugin
npm install
npm test          # the node:test (tsx --test) suite — the local gate
npm run bundle    # esbuild -> out/extension.js + out/uninstall.js (CJS)
npm run package   # npx @vscode/vsce package -> a single .vsix

The bundle inlines the shared IPC client and marks vscode external, so the .vsix is self-contained and ships no repo source (see .vscodeignore).

Publishing (maintainers)

Publishing is a deliberate, manually-triggered action — never a per-push CI job — and targets the VS Code Marketplace only:

Set the publisher field in package.json to your Marketplace publisher id (currently insors).
Add a Marketplace Personal Access Token as the VSCE_PAT GitHub Actions secret (it lives only as a secret — never in the repo).
Bump the version in package.json.
Run the vscode-plugin workflow from the GitHub Actions tab (workflow_dispatch). It bundles, packages, and vsce publishes to the Marketplace.

License

MIT — see LICENSE.

insrc

insors