Heap Code
Model-agnostic AI coding assistant for VS Code. Works with any OpenAI-compatible API — local (Ollama, LM Studio, vLLM, LocalAI) or cloud (OpenAI, OpenRouter, Groq, Together, Azure OpenAI, NVIDIA NIM).
Privacy
Your code never leaves your machine except to the model endpoint you configure — no proprietary backend, no account. Point it at Ollama on your LAN and your code stays home. API keys live in your OS keychain, never in settings files.
Telemetry
Heap Code sends anonymous usage telemetry by default: which commands/features are used and coarse error counts, tagged with a random per-install ID (not your identity), extension/VS Code version, and OS. It never includes your code, prompts, file contents, or file paths.
To turn it off:
- Extension setting:
heapcode.telemetry.enabled → false, or
- VS Code's own
telemetry.telemetryLevel → off (this disables it for every extension, including Heap Code, regardless of the setting above).
Quick start
- Open the Heap Code icon in the activity bar.
- Click the ⚙ gear in the chat header → add a provider profile (Ollama works out of the box at
http://localhost:11434/v1).
- Pick a model, chat. Switch to Agent mode for autonomous multi-file tasks.
Features
- Chat — streaming markdown; history; slash commands (
/explain, /fix, /review, /security-review…); @selection @file @problems @terminal @workspace mentions; attach files, folders, or a line-range selection; paste screenshots for vision models; code-block Copy / Insert / Apply
- Agent mode — plans, then reads, searches, edits files and runs commands autonomously; language-server tools (symbols, references, definitions) plus a persisted
repo_map outline of the workspace; every non-read action behind permission prompts; per-file Keep / Revert / Reapply and per-turn workspace restore; a 🔧 picker chooses which tools it may use; proposes short "worth remembering" notes at the end of a session, saved only with your confirmation
- Completions — ghost text with proper FIM formats per model family (Qwen, DeepSeek, StarCoder, CodeLlama, Codestral, CodeGemma), native Ollama FIM, debounced + cancellable; repo-level context from the semantic index (fast keyword-only while typing, full semantic search on manual trigger)
- Inline edit — select code,
Ctrl+I/Cmd+I, describe the change, review in a native diff, accept from its title bar or the Accept/Reject CodeLens above the diff
- Semantic search (RAG) — background incremental index with your embeddings model; AST-aware chunking (TypeScript/TSX/JavaScript/JSX/Python) so results align to real function/class boundaries; hybrid retrieval (embeddings + BM25 keyword search, fused) on by default; optional contextual retrieval for even better recall; powers
@workspace and the agent's search
- Context management — usage meter with auto-detected model context windows; automatic conversation compaction; edit any earlier prompt to rewind conversation and workspace
- MCP — register Model Context Protocol servers (stdio / HTTP / SSE); their tools join agent mode under the same permissions
- Project memory —
HEAPCODE.md (or AGENTS.md, if that's what your project already uses) + .heapcode/memory.md load into every session
- Git — commit-message generation from the staged diff (✨ in Source Control)
Provider profiles
Named profiles with per-role models — chat, edit, apply, autocomplete, agent, embeddings, rerank, context — managed in the in-chat settings panel or via the status bar. Capabilities (vision, tool calling, context window) resolve per preset and per model, with per-profile overrides.
License
PolyForm Noncommercial License 1.0.0 — free for any noncommercial use (personal, private/internal, research, nonprofit, education, government); not for commercial productization. 0.1.x remains available under its original Apache-2.0 terms. Source: github.com/heaplabshq/heapcode
| |
