
Lattice AI — VS Code Extension
Extension for the local-first Digital Brain that keeps your knowledge durable across any AI model.

Lattice AI connects VS Code, Cursor, and VSCodium to your local-first Digital
Brain. Use it to send files into durable Brain context, work with Brain-backed
chat, run model workflows, and trigger coding actions from the editor.
Current Release
13.4.0 — Living Synthesis introduces the model-agnostic Universal Adaptive Agent Harness (probe.rs 3-turn micro-probing, observation-first ranking & bilingual domain classification, deterministic UI bounding-box center resolution ui_resolve.rs, malformed-JSON repair & ChatML prefill, active vs. superseded evidence separation rag_grounding.rs), modular native web fetching & Computer Use bridge parity, a reproducible 25-task live empirical benchmark (Qwen2.5-0.5B-Instruct-Q4_K_M 28.0% → 84.0%, +56.0 pp), and 100% EN ↔ KO bilingual documentation parity alongside 13.1.0–13.3.1 capabilities (two-phase async semantic incremental chunk sync, claim-verification & epistemic cards in AnswerProof.tsx / BrainIntelligencePanel.tsx, interactive Knowledge Graph neighborhood & 2-hop synthesis, contradiction inspector, Chat Save-to-Brain & inline proposals, live index drain, cooperative download cancel, Command Palette KG focus, and reactive workspace switching at 100% Vitest coverage). Full disclosure is in docs/releases/RELEASE_NOTES_v13.4.0.md.
13.0.0 — Sovereign Synthesis enforces cross-domain decoupling (lattice_core::redact, lattice_agent::seam, domain_boundaries.rs) and unified Rust server i18n (lattice_core::messages), adds 2-Hop Knowledge Graph reasoning chains, bridge concepts & contradiction synthesis (reasoning_paths.rs, MultiHopSynthesisView.tsx), introduces single-transaction review store mutations, one-click reversible folder reorganization undo (ReviewCard.tsx) and durable conversation document sessions across restarts, and splits frontend/src/api/client.ts and src-tauri/src/backend.rs at 100% Vitest coverage. Full disclosure is in docs/releases/RELEASE_NOTES_v13.0.0.md.
12.10.0 — Grounded Self adds deterministic agent transcript grounding & interactive citation proof badges, a Brain Self-Model panel with inline proposal approval/dismissal, Knowledge Graph PART_OF edge cascade parity across edges and edges_v2, cooperative 499 model download cancellation with automatic post-download semantic embedding transition & reindex drain, crate-internal http_reply.rs transport decoupling, 100% Python worker server-i18n lock (6/6 routers), parallel cargo nextest execution, and chained release.yml → publish.yml CD publishing. Full disclosure is in docs/releases/RELEASE_NOTES_v12.10.0.md.
12.9.0 — Semantic Guard aligns query-time semantic embeddings with the active EmbeddingIndexContract across search/recall/chat, adds pure-Rust NativeHnswIndex & workspace/model-scoped Python HNSW sidecars, enforces fail-closed local_only/cloud_allowed network boundary & cloud context guards, hardens Tauri desktop / LTCAI CLI / host supervisor / multi-stage Docker lifecycles, streams live Workspace OS SSE events (/api/realtime/stream), and upgrades v2 .latticebrain archive KDF & CI gates. Full disclosure is in docs/releases/RELEASE_NOTES_v12.9.0.md.
12.8.0 — Trusted Brain unifies the EmbeddingIndexContract & knowledge lifecycle tracking, pushes workspace/lifecycle filtering before SQL LIMIT with aligned vector search & bounded sidecar caches, enforces fail-closed multi-user/workspace/tenant isolation across gateway routes, delivers AEAD-encrypted .latticebrain archive portability with honest planned PostgreSQL status, modularizes Library/System/Capture UI slices, and automates Apple Silicon DMG CI/CD releases. Full disclosure is in docs/releases/RELEASE_NOTES_v12.8.0.md.
12.7.0 — Fast Index adds a Pre-Tokenized BM25F Sidecar (Bm25SidecarIndex), SQLite bm25_node_stats persistence, and an inverted posting-list & GraphRAG adjacency cache for O(1) lookup without query-time re-tokenization, alongside a product-first README.md and v12.0.0+ release-history polish. Full disclosure is in docs/releases/RELEASE_NOTES_v12.7.0.md.
12.6.0 — Hybrid Graph adds 3-way BM25F + Vector + Knowledge Graph RRF hybrid indexing & graph-aware agent skills, multi-scale subword FallbackTextEmbedder, unauthenticated EmbeddingGemma 2 public mirrors & quantized GGUF/MLX/ONNX fallbacks, real-time SSE download progress streaming (MB/s & ETA) in Onboarding and Model Library, Pillow 13+ cleanup, Ubuntu 24.04/26.04 t64 CI resilience, and the modular scripts/bump_version/ CLI. Full disclosure is in docs/releases/RELEASE_NOTES_v12.6.0.md.
12.5.0 — Deep Vector adds native Google EmbeddingGemma 2 support, hardware-aware PC-tailored embedding recommendations, Pillow 13+ flattened data compatibility, Node 24 native CI/CD, and automated OpenAPI specification and fixture synchronization. Full disclosure is in docs/releases/RELEASE_NOTES_v12.5.0.md.
12.4.0 — Sharp Weight keeps the One Door topology (422 operations / 41 families, 20-route worker), adds verified mlx-community/Qwen3.8-27B-4bit, mlx-community/gemma-4-26B-A4B-it-qat-4bit, and mlx-community/gemma-4-26B-A4B-it-qat-OptiQ-4bit local/cloud routes, and closes all Python (pillow>=13,<14) and npm dependency security advisories. Full disclosure is in docs/releases/RELEASE_NOTES_v12.4.0.md.
12.3.0 — Live Wire keeps the One Door topology (422 operations / 41 families, 20-route worker) and verifies Hugging Face GGUF quantization tags (Q4_0 for Gemma 4, MXFP4 for GPT-OSS 20B), Qwen3.5-9B / Qwen3.6-35B-A3B GGUF routes, auxiliary GGUF filtering + companion --mmproj download, small-model tool/MCP argument recovery, and native HTML structure inspection (/tools/inspect_html) + RFC 5987 UTF-8 download (/tools/download). Full disclosure is in docs/releases/RELEASE_NOTES_v12.3.0.md.
12.2.1 — True Count keeps the One Door topology (422 operations / 41 families, 20-route worker) and makes search, skip, watch, and small-model summaries count what is actually there. Full disclosure is in docs/releases/RELEASE_NOTES_v12.2.1.md.
The extension surface is unchanged apart from its version. Sync, status, recall, staged proposals, POST /agent with stream:true, and strict / trusted / bypass travel over the wire as before.
12.0.0 — Open House grouped the two largest crates by purpose (lattice-agent six groups, lattice-platform seven domains, all git mv) and closed four named gaps: restore is live without a restart, /setup/install installs on per-item consent, POST /mcp is inside the OpenAPI contract, and pointer tools are pip install "ltcai[pointer]". Full disclosure of that release is in docs/releases/RELEASE_NOTES_v12.0.0.md.
The extension surface is unchanged apart from its version. Sync, status, recall, staged proposals, POST /agent with stream:true, and strict / trusted / bypass travel over the wire as before — the crate regrouping renamed files, not routes. Cloud escalation and MCP live on the host, not in the editor.
The extension connects your editor to the same local-first Digital Brain, explicit consent gates, replaceable model workflow (MLX-VLM / llama.cpp / Ollama / LM Studio / optional cloud), and proposal-first Review Center used by the desktop and web app.
Quick Start
Install and start the Lattice AI workspace:
pip install ltcai
LTCAI
For Apple Silicon local model support:
pip install "ltcai[local]"
Then install the extension:
The extension auto-connects to http://localhost:4825.
What It Adds
| Area |
Description |
| Chat panel |
Work with local or cloud models using workspace and graph context |
| Current file actions |
Ask, edit, explain, refactor, and generate tests from the active file |
| Send To Lattice |
Send the current file or selection into workspace workflows |
| Sync status |
See whether the editor is connected, indexing, synced, or offline |
| Model workflows |
Load, switch, and use model choices managed by Lattice AI |
| Knowledge Graph |
Use graph-backed context from files, notes, screenshots, and conversations |
| Multi-agent workflow |
Start planner/executor/reviewer style coding workflows |
| Local-first workspace |
Keep personal work local while supporting organization workspace patterns |
| Native v4 app |
Use /app as the primary product shell |
Commands
| Command |
Shortcut |
Description |
| Lattice AI: Open Chat |
Cmd+Shift+A |
Open the workspace chat panel |
| Lattice AI: Edit Selection |
Cmd+Shift+E |
Rewrite selected code |
| Lattice AI: Load Model |
Cmd+Shift+M |
Pick a local or cloud model |
| Lattice AI: Explain Selection |
Right-click menu |
Explain selected code |
| Lattice AI: Refactor Selection |
Command palette |
Refactor selected code |
| Lattice AI: Generate Tests |
Command palette |
Generate focused tests |
| Lattice AI: Send To Lattice |
Command palette |
Record file context in Lattice AI |
| Lattice AI: Ask About Current File |
Command palette |
Ask with current file context |
| Lattice AI: Show Sync Status |
Command palette / status bar |
Inspect the editor-to-app bridge state |
| Lattice AI: Generate Terminal Command |
Right-click menu |
Generate a shell command |
| Lattice AI: Save to Knowledge Garden |
Right-click menu |
Save a snippet or note |
Model Workflow
Lattice AI supports local and cloud model choices:
- MLX-VLM on Apple Silicon for current multimodal local models, with MLX-LM
retained as a text fallback only for standard Gemma 4 metadata.
- LM Studio, vLLM, llama.cpp, and Ollama-compatible local paths.
- OpenAI, OpenRouter, Groq, Together AI, and OpenAI-compatible endpoints.
- Model cards disclose maker country, maker company, run mode, internet usage,
model name, HF verification status, download/load strategy, and hardware fit.
Settings
| Setting |
Default |
Description |
ltcai.serverUrl |
http://localhost:4825 |
Lattice AI workspace URL |
ltcai.autoLoadModel |
false |
Load the default model on extension activation |
ltcai.defaultModel |
empty |
Optional default model id |
Remote or tunnel server example:
{
"ltcai.serverUrl": "https://your-server.example"
}
Links
License
MIT — TaeSoo Park