Skip to content
| Marketplace
Sign in
Visual Studio Code>Other>AIRK — Artifact Indexing and Retrieval Kernel (beta)New to Visual Studio Code? Get it now.
AIRK — Artifact Indexing and Retrieval Kernel (beta)

AIRK — Artifact Indexing and Retrieval Kernel (beta)

Ioka LLC

| (0) | Free
Pre-release beta: local collection indexing and specialized MCP evidence for coding agents, with a bundled native helper.
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

AIRK for VS Code — Beta 0.1.0

AIRK supplies local code and document evidence to coding agents through fourteen read-only MCP tools. This preview bundles its native helper; installation requires no Docker, Rust, Python, service account or API key.

Your agent can find a definition, inspect nearby source lines or retrieve a document requirement through a shared workspace index. AIRK returns bounded evidence with source paths, hashes and original locations, so you can inspect what the agent used. The agent supplies the reasoning; smaller responses do not by themselves prove lower bills or better answers.

Workspace code and documents flow through AIRK's local collection to independently configured Claude Code, Codex, Gemini CLI and other MCP clients.

Claude Code, Codex and Gemini CLI: choose AIRK: Set Up Coding Agent to preview the selected client's configuration. Claude Code uses a project .mcp.json entry; Codex uses a mcp_servers TOML table; Gemini CLI uses mcpServers in .gemini/settings.json. Merge the scoped entry, then enable it in that client. Several agents can share the same root, data directory and collection. VS Code's MCP registration does not configure the other clients automatically.

Saved files are indexed into a complete published revision, then an agent retrieves bounded evidence. Coverage, controls and citations make freshness and scope visible.

These are explanatory illustrations. AIRK's local runtime does not control whether your selected agent sends returned evidence to its model provider.

Document facts include heading context, dates, requirement IDs and modality, numeric bounds/units, qualitative cues and design/architecture/feature/decision candidates. DOCX tables and XLSX worksheets retain structured row/cell evidence with original locations; stored spreadsheet formulas are never recalculated. Use Index Now after updating to populate new facts in existing collections. Categories are evidence cues, not inferred compliance; ambiguous dates and source locations remain explicit.

Extension identity: iokaio.airk. Display name: AIRK — Artifact Indexing and Retrieval Kernel (beta). Version 0.1.0 has platform-specific packages; check the package's preview/stable channel before installing. The Windows package is airk-0.1.0-win32-x64.vsix; Linux and macOS use the matching platform suffix.

Install the VSIX for your platform through Extensions: Install from VSIX. Open a folder and grant workspace trust. The AIRK status item shows measured coverage for the active folder; click it to index. Use AIRK: Open Controls in the command palette for other controls. Each workspace folder gets an isolated collection. VS Code discovers the collection through AIRK's MCP provider; use MCP: List Servers to enable it.

  • Index Now / Reindex Collection: incremental indexing or full re-extraction.
  • Choose Indexing Mode: after edits settle (default), when idle, on save, or manual.
  • Pause / Resume Auto-indexing: preserve queued edits; an in-progress write finishes.
  • Choose Index Storage Folder: put indexes on a drive of your choice. Old indexes remain in their old location; this changes storage without migrating or deleting them.
  • Manage Collections and Disk Space: inspect collection bytes, available drive space, known roots, last indexed time and generation. Remove an unused collection through the separate removal command; source files remain intact.
  • Trim Obsolete Index Generations: review reclaimable generations, then confirm native pruning; active evidence and reader leases remain protected.
  • Show Collection Status: inspect current, stale, missing, extraction-failed and capacity-blocked eligible files. Incomplete scans show unknown coverage; unsaved buffers are separate from indexed disk bytes.
  • Show Local Usage and Estimated Savings: inspect content-free local aggregates and signed context estimates. These are not measured provider tokens or billed savings.
  • Set Up Coding Agent: preview a scoped configuration for Claude Code, Codex, Cursor, Gemini CLI, Cline, Roo, Junie, Copilot CLI, OpenCode or Kilo. Amazon Q and Windsurf/Devin get connection details and version-specific setup guidance. AIRK never overwrites their settings or enables blanket tool approval.
  • Verify Native MCP Connection: check the actual native handshake, fourteen tools and collection status. This does not prove another agent has enabled the server.

airk.storage.path and airk.storage.maxBytes are machine settings. The optional byte budget pauses automatic indexing after existing collection usage reaches the limit; manual commands may exceed it. It is a soft limit, not a filesystem quota. No collections are deleted automatically. Changing storage requires reconnecting already-running agents. Helper paths cached per version/digest survive extension directory replacement; refresh agent configuration after helper upgrades.

Filesystem events coalesce in the editor; submitted batches enter a durable native queue shared by hosts using the same data root. Pending/completed revisions, pause and failure state survive helper exits. Repeated clicks join the active pass; later changes remain queued. On-save mode does not automatically index external agent edits. Idle mode waits for editor activity as well as filesystem quiet. Coverage scans run separately after changes, after indexing and periodically; they do not restart MCP clients.

Retrieval uses persistent Tantivy lexical indexes, SQLite exact facts/relationships and optional Microsoft DiskANN search through AIRK's disk file provider. Deliberately supplied, digest-checked local word-vector assets enable vector/hybrid queries; no model is downloaded and no remote fallback is used. This initial model adapter is not transformer inference and its production semantic quality remains unqualified. Code evidence includes scopes/signatures/call/framework candidates; optional producer-bound SCIP imports add reference/implementation assertions. Use Index Now after upgrading to migrate older state.

The native helper streams artifacts into content-addressed storage and supports source aliases, shared logical-byte quotas and explicit verified data-root relocation through the CLI. The editor's storage picker retains its existing selection behavior. OCR, live LSP/Ctags and large-corpus qualification remain open. Safety limits are 1 MiB per source, 64 MiB eligible source, 128 MiB extracted text per collection and 10,000 files. Skipped-file details appear in AIRK Output. Index contents and aggregate usage stay local; the extension has no runtime telemetry or network service. Your selected agent controls any subsequent provider requests.

Packages target Windows x64, Linux x64 and both Mac architectures. Remote workspaces run the extension/helper on their workspace host and require a matching platform build; remote editor and oldest-OS qualification remain open. Browser-only workspaces cannot run this native helper. Signing, notarization and Marketplace publication are separate release work.

Source, local build instructions and support limits: AIRK repository, extension documentation. Licensed under Apache-2.0; bundled dependency notices retain their respective terms.

  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft