AGMark
Interactive markdown annotation system for AI agent collaboration. Select text in VSCode to annotate — Claude processes annotations in the same session via MCP tools.
Workflow
Select text → Annotate → Send to Claude → MCP processes → Edit document → clean
───────────────────────────────────────────────────────────────────────────────→
One review cycle
Quick Start
VSCode Extension
cd vscode-extension
npm install
npm run build
Right-click any .md file → AGMark Preview, or use the command palette: AGMark: Init Annotations.
Rust CLI
cargo build --release
./target/release/agmark --help
MCP Server
Register in your Claude Code configuration:
{
"mcpServers": {
"agentmark": {
"command": "node",
"args": ["path/to/vscode-extension/mcp-server.js"]
}
}
}
Project Structure
agmark/
├── crates/ # Rust workspace
│ ├── core/ # Annotation storage, anchor resolution, engine
│ ├── cli/ # agmark CLI (10 subcommands)
│ └── tui/ # agmark-tui terminal browser
├── vscode-extension/ # VSCode extension
│ ├── src/
│ │ ├── webview/ # Preact UI + highlight engine
│ │ ├── __tests__/ # Unit tests (vitest + jsdom)
│ │ ├── AnchorResolver.ts # Three-level anchor resolution
│ │ └── CommentableEditor.ts
│ └── mcp-server.js # MCP server (stdio)
├── docs/ # Design documentation
└── README.md
Anchor Strategy
Three-level progressive resolution to survive document edits:
- Heading path + paragraph index — 100% accurate when structure is unchanged
- Fuzzy text fingerprint matching — prefix edit distance + Jaccard + trigram Dice
- Orphan — degraded marker: "paragraph changed"
| Tool |
Description |
list_pending |
List all open threads across the project |
get_annotations(document, status?) |
Get threads for a document |
reply_to_annotation(document, threadId, body, resolve?) |
Reply and resolve a thread |
refresh_document(document) |
Signal the extension to refresh after edits |
get_stats(document?) |
Annotation statistics |
Tests
cd vscode-extension
npm test # vitest run (79 tests, terminal-only)
npm run test:watch
Documentation
