A VSCode extension that makes every piece of Claude Code's context — skills, agents, hooks, instructions, MCP servers, plugins, and memory — visible, attributable, manageable, and observable, directly from the editor.
Working term: AI artifacts — any file or config entry that shapes what a Claude Code session knows or can do. This extension is an AI-artifact manager the way the SCM view is a git manager.
Features
- Context Explorer — every AI artifact in play, grouped by type or by scope (Managed / User / Project / Local / Plugin / Directory-scoped), with shadowing warnings, disabled-state badges, and live session-usage highlights.
- Provenance & inspection — a detail view per artifact: rendered content, frontmatter, token weight, contributing file, and the merged effective settings with per-key provenance.
- Creation & management — create skills/agents/rules/hooks/output-styles from templates,
move artifacts between user and project scope, disable without deleting (
.disabledrename), delete with confirmation. Safety rails warn when a live session won't see your change. - Discover — browse and install from the official Anthropic plugin marketplaces, the official MCP registry (REST), community aggregators (deep-link + paste-to-install), and any custom git marketplace (GitLab supported). Every install ends in a review step showing exactly what will be written and flagging executable surface — never one silent click.
- MCP management — all configured servers with scope and transport, duplicate-name
shadowing visualised, add/edit via form or raw JSON (writes through
claude mcp add/removewhen the CLI is present), connection status viaclaude mcp list. - For This File — the exact instruction context Claude assembles for the active editor
file: the CLAUDE.md chain with
@importexpansion,paths:-matched rules, and directory-scoped skills. Status-bar count + optional CodeLens. - Session awareness — passive transcript watching (zero footprint in your Claude config):
artifacts light up as the live session loads/uses them, and a Session Dashboard shows
artifacts touched, MCP call counts, subagents spawned, and running true API cost
computed from per-message
usagetoken fields × a model-pricing table. - Health — duplicates/shadowing, secrets in committed config, machine-specific paths,
dead weight (unused across tracked sessions), oversized always-loaded instructions, broken
@imports/[[memory-links]]/hook scripts, deprecated transports — plus a context-budget bar (bundled-BPE token estimates; exact counts via the Anthropiccount_tokensAPI when an API key is configured).
Requirements
- VSCode ≥ 1.96.
- Claude Code installed (
claudeCLI optional — everything degrades to file-level operation without it; plugin installs and MCP status need it).
Settings
| Setting | Default | Purpose |
|---|---|---|
claudeContext.sessionWatch.enabled |
true |
Watch session transcripts for live usage |
claudeContext.marketplaces.custom |
[] |
Git URLs of custom marketplaces / skills repos |
claudeContext.marketplaces.aggregators |
all | Community aggregator deep-link sources |
claudeContext.tokens.anthropicApiKeyEnvVar |
ANTHROPIC_API_KEY |
Env var for exact token counts |
claudeContext.tokens.countModel |
claude-sonnet-5 |
Model for count_tokens calls |
claudeContext.fileLens.codeLens |
false |
CodeLens line at the top of files |
claudeContext.claudeDir |
~/.claude |
Override for testing |
A note on transcript watching
Session awareness reads the JSONL transcripts Claude Code already writes under
~/.claude/projects/<slug>/. That format is documented as internal and may change; the parser
degrades per-line (a format break turns off session highlights, never the rest of the
extension). Nothing is ever installed into your Claude configuration.
Development
npm install
npm run build # extension (esbuild) + webviews (vite)
npm test # vitest unit suite (fixture corpus in fixtures/)
npm run typecheck
npm run lint
npm run package # .vsix via vsce
Architecture and the full feature specification live in PLAN.md.