AgentCanvas

View your coding agent's structured output in VS Code — tables, Mermaid diagrams, charts, math, and prose, all rendered in a panel.

Documents are written by your own coding agent (Claude Code / Codex / Gemini CLI) running a bundled Agent Skill, under your own login. The extension itself is a pure renderer: it reads JSON from disk and draws it. It never calls a model, touches OAuth/API keys, launches an agent, or makes a network call.

Why AgentCanvas

• Use the subscription you already have. Docs come from your existing Claude / Codex / Gemini agent — no extra subscription, no API key, no per-token billing.
• Fully local & private. Zero model or network calls from the extension. No telemetry, no accounts. Output files never leave your machine.
• Rich rendering. KaTeX math, Mermaid diagrams (architecture flowcharts, sequence, ER, pie, xychart), declarative scatter/line charts with log axes and error bars, code blocks, tables, and prose.
• Read like a doc. Highlight passages and attach notes (kept locally, searchable from the Notes panel, surviving doc regeneration), and Copy^p any selection with its file path and section for pasting back to your agent with provenance.

How it works

your agent runs the agentcanvas skill (reads context, writes JSON)
        │
        ▼
.agentcanvas/docs/<docId>.doc.json   ◄── validated by schema/document.schema.v1.json
        │
        ▼
AgentCanvas renders it in the viewer (live-reloads on change)

Requirements

AgentCanvas is a renderer, not an AI client. To produce docs you need a coding agent installed locally — Claude Code, Codex, or Gemini CLI — running under your own login.

Getting started

1. Install the skill. Run AgentCanvas: Install Skill from the Command Palette. It copies agentcanvas into .agents/skills/, .claude/skills/, and .gemini/skills/ (it only writes files — it never launches an agent).
2. Open the viewer. Run AgentCanvas: Open from the Command Palette.
3. Ask your agent to run agentcanvas — e.g. "make a comparison table for X vs Y", "draw an architecture diagram of this codebase", "document the DB schema as an ER diagram", "draw a sequence diagram of the login flow".
4. The doc appears in the Docs dropdown and renders live.

Commands

Privacy & compliance

Everything is local. The extension makes zero model/network calls — inference happens only inside your own agent, under your own login. No telemetry, no accounts, no API keys.

License

MIT