feat: on-demand context for Claude Code via MCP server (v0.5.0)

.eslintrc.cjs

@@ -23,5 +23,5 @@ module.exports = {
     '@typescript-eslint/no-unused-vars': ['warn', { argsIgnorePattern: '^_' }],
     '@typescript-eslint/no-explicit-any': 'warn',
   },
-  ignorePatterns: ['dist', 'node_modules', '*.config.js', '*.config.ts'],
+  ignorePatterns: ['dist', 'node_modules', '*.config.js', '*.config.ts', 'mcp-server'],
 };

README.md

@@ -56,6 +56,9 @@ I see the same.
 - **CLAUDE.md への直書き** *(v0.2.0+)*：プロジェクトの `CLAUDE.md` を一度ピックすれば、以降のキャプチャは File System Access API 経由で自動 append。コピペ不要
 - **マルチプロジェクト振り分け** *(v0.3.0+)*：URL パターン（glob）で複数の CLAUDE.md を使い分け。例：`github.com/anthropic/*` → anthropic 用、`zenn.dev/*` → 個人メモ、無マッチは default ルート
 - **claude.ai 会話のキャプチャ** *(v0.4.0+)*：claude.ai でブレストした会話をワンクリックで CLAUDE.md に流し込み。share リンクの 403 問題、手コピペの労力を解消。内部 API を叩いて thinking blocks・tool_use・branch 構造まで保持
+- **アーティファクト抽出 & 範囲選択** *(v0.5.0+)*：claude.ai キャプチャで「**アーティファクトだけ**」（Claude が書いたコード／ドキュメントを整形済みコードブロックで抽出、会話は捨てる）や「**直近 N 件のみ**」を選択可能。長い設計会話から必要な部分だけを取り込めます。アーティファクトは同一 ID の更新を最新版に集約。claude.ai のチャットを開いているときは**ポップアップから直接トグル**できます
+- **同一会話は上書き更新（siloを作らない）** *(v0.5.0+)*：MCP ストアモードで同じ claude.ai 会話を再キャプチャすると、**新しいスナップショットを増やさず既存ファイルを上書き**します。会話タイトルが変わっても安定 ID で追跡するので、ストアは常に single source of truth に保たれます（古いアーティファクトファイルも掃除）
+- **MCP 経由でオンデマンド参照** *(v0.5.0+)*：キャプチャを `CLAUDE.md` に追記する代わりに、専用ディレクトリへ 1 件 1 ファイルで保存し、付属の [MCP サーバ](./mcp-server) が Claude Code に**必要なときだけ**渡します。Anthropic 公式の「`CLAUDE.md` は小さく保て（肥大化すると指示が無視される）」ガイドと整合し、コンテキストを汚しません。`list_contexts` / `search_contexts` / `get_context` で検索・取得できます。→ [セットアップ](./mcp-server/README.md)
 - **キャプチャバッファ**：複数ページをまとめて溜めて、後から一括エクスポート
 - **キーボードショートカット**：
   - `Ctrl+Shift+L`（macOS: `Cmd+Shift+L`）：ページ全体
@@ -257,6 +260,9 @@ In short: a clipper purpose-built for AI agent context files. If you want a gene
 - **Direct CLAUDE.md write** *(v0.2.0+)* — Link a `CLAUDE.md` once via the File System Access API; subsequent captures append directly, no copy/paste
 - **Multi-project routing** *(v0.3.0+)* — Link multiple `CLAUDE.md` files with URL glob patterns. Captures from `github.com/anthropic/*` go to one file, `zenn.dev/*` to another, unmatched URLs to a default route
 - **claude.ai conversation capture** *(v0.4.0+)* — Capture your claude.ai brainstorm conversation in one click. Bypasses the 403-on-share-link issue and the manual copy-paste loop. Hits claude.ai's internal API to preserve thinking blocks, tool_use entries, and branch structure that DOM scraping would miss
+- **Artifact extraction & range selection** *(v0.5.0+)* — For claude.ai captures, choose **artifacts only** (extract just the code/documents Claude wrote as clean fenced code blocks, dropping the conversation) or **keep the last N messages** (trim a long planning thread to what matters). Artifacts are deduped by id so an updated artifact is captured at its final version, not every intermediate edit. When you're on a claude.ai chat, these are **toggleable right from the popup**
+- **Same conversation updates in place (no silos)** *(v0.5.0+)* — In MCP-store mode, re-capturing the same claude.ai conversation **overwrites the existing file instead of piling up new snapshots**. It's tracked by a stable id, so even if the conversation's title changes the store stays a single source of truth (stale artifact files are cleaned up too)
+- **On-demand access via MCP** *(v0.5.0+)* — Instead of appending to `CLAUDE.md`, save each capture as a standalone file in a directory, and let the bundled [MCP server](./mcp-server) hand them to Claude Code **only when it asks**. This aligns with Anthropic's own guidance to keep `CLAUDE.md` small (bloated context files make Claude ignore your instructions) — your research stays available without polluting context. Searchable via `list_contexts` / `search_contexts` / `get_context`. → [Setup](./mcp-server/README.md)
 - **Buffer mode** — Stack multiple captures and export all at once
 - **Keyboard shortcuts**:
   - `Ctrl+Shift+L` (Cmd+Shift+L on macOS) — capture page

docs/issue-13843-comment.md

@@ -0,0 +1,35 @@
+# Draft comment for anthropics/claude-code#13843
+
+> "Share conversation context from Claude.ai to Claude Code"
+> https://github.com/anthropics/claude-code/issues/13843
+
+---
+
+For anyone who wants this **today** without waiting on a first-party fix, I built an open-source pair of tools that bridges it — and deliberately avoids the native-messaging-host setup that trips up most attempts at this.
+
+**How it works:**
+1. A Chrome extension captures your claude.ai conversation — preserving thinking blocks, `tool_use` steps, and branch structure (it reads claude.ai's conversation API rather than scraping the rendered DOM, so artifacts and code survive).
+2. It saves each conversation as a Markdown file in a directory you pick (via the File System Access API — **no native messaging host**, which is where similar tools usually break).
+3. A small MCP server exposes those files to Claude Code on demand: `list_contexts`, `search_contexts`, `get_context`.
+
+**Why on-demand instead of dumping into `CLAUDE.md`:** Anthropic's own [memory guidance](https://code.claude.com/docs/en/memory) is to keep `CLAUDE.md` small — bloated context files make Claude ignore instructions. So the agent pulls in a captured conversation only when it needs it, instead of permanently inflating context. (No silos either — it's a single directory you can inspect, search, or delete from.)
+
+Setup is `npx` + one line:
+
+```bash
+claude mcp add ccc-contexts -- npx -y claude-code-context-capturer-mcp ./.ccc-contexts
+```
+
+Repo: https://github.com/OceansCreative/claude-code-context-capturer
+It also captures general web pages (GitHub issues, docs, Stack Overflow, etc.) the same way, if that's useful for research → code workflows.
+
+Feedback very welcome — especially on whether on-demand MCP access fits how you actually move from planning in claude.ai to building in Claude Code.
+
+---
+
+## Posting notes (for me, not part of the comment)
+
+- **Don't post until** the MCP server is actually published to npm (the `npx` line must work for a stranger) OR reword to point at the manual install. Posting a broken install command is exactly what sank the competitor.
+- Tone: helpful, not spammy. Lead with "here's a thing that works today," not "check out my project."
+- The thread has a vocal "no export/import, I want live read access / single source of truth" camp (powerobject, +10). The "single directory you can inspect/search/delete, agent reads on demand" framing is aimed at them — keep it.
+- Consider also replying to geovaniprodata, who hit "native messaging host not found" with the competitor — explicitly note this design has no native host.

mcp-server/README.md

@@ -0,0 +1,179 @@
+# Claude Code Context Capturer — MCP server
+
+Expose the web pages and **claude.ai conversations** you capture with the
+[Claude Code Context Capturer](https://github.com/OceansCreative/claude-code-context-capturer)
+browser extension to **Claude Code — on demand**, over MCP.
+
+## Why on-demand instead of CLAUDE.md?
+
+Anthropic's own guidance is to keep `CLAUDE.md` small (target < 200 lines) —
+[bloated context files make Claude ignore your instructions](https://code.claude.com/docs/en/memory).
+Dumping every captured article and chat into `CLAUDE.md` works against that.
+
+This MCP server takes the opposite approach: your captures live as individual
+Markdown files in a directory, and Claude Code pulls in **only the ones it needs,
+only when it needs them** — via `list_contexts`, `search_contexts`, and
+`get_context`. Nothing is loaded into context until the agent asks. No silos,
+no bloat.
+
+## How it fits together
+
+```
+ ┌─────────────┐   capture (FSA write)    ┌──────────────────┐   read    ┌─────────────┐
+ │  browser    │ ───────────────────────▶ │  contexts dir/   │ ◀──────── │ this MCP    │
+ │  extension  │   <slug>.md per capture   │  *.md files      │           │ server      │
+ └─────────────┘                           └──────────────────┘           └──────┬──────┘
+                                                                                  │ stdio (MCP)
+                                                                           ┌──────▼──────┐
+                                                                           │ Claude Code │
+                                                                           └─────────────┘
+```
+
+**No native-messaging host.** The extension writes plain files via the File
+System Access API; this server just reads that directory. That's the single
+biggest reliability difference from similar bridges — there's nothing fragile to
+register.
+
+## Install
+
+Requires Node.js ≥ 18. No global install needed — `npx` fetches it on demand.
+
+```bash
+# From inside your project, register the server pointed at your captures dir:
+claude mcp add ccc-contexts \
+  -- npx -y claude-code-context-capturer-mcp ./.ccc-contexts
+```
+
+Or add it to your MCP config manually:
+
+```json
+{
+  "mcpServers": {
+    "ccc-contexts": {
+      "command": "npx",
+      "args": ["-y", "claude-code-context-capturer-mcp", "./.ccc-contexts"]
+    }
+  }
+}
+```
+
+The captures directory is resolved in this order:
+
+1. the first CLI argument (`./.ccc-contexts` above)
+2. the `CCC_CONTEXTS_DIR` environment variable
+3. `<cwd>/.ccc-contexts`
+
+`~` is expanded to your home directory.
+
+## Connect the extension
+
+1. Open the extension's options page.
+2. Under **MCP contexts directory**, click **Link directory** and pick the same
+   directory you pointed the server at (e.g. `.ccc-contexts` in your project).
+3. Set **Default action** to **“Save to MCP contexts directory.”**
+
+Now every page or claude.ai conversation you capture lands as a `.md` file the
+server can serve to Claude Code.
+
+## Tools
+
+| Tool | Description |
+|------|-------------|
+| `list_contexts` | List captures (metadata only). Filter by `parser`, `tags` (AND/OR via `tagMatch`), `since`/`until` date range; `limit`. |
+| `search_contexts` | Ranked keyword search across title, URL, tags, and body, with snippets. Accepts the same `parser`/`tags`/`since`/`until` filters to scope the search first. |
+| `stats_contexts` | Aggregate overview: total count + size, breakdown by source, top tags, captured-at date range. |
+| `get_context` | Fetch one capture by `slug` (fuzzy matching). `format:"code"` returns just the raw code (fences + frontmatter stripped) — ideal for single artifacts. |
+| `delete_context` | Remove a capture by `slug` once it's been incorporated. |
+
+### Artifacts as individual files
+
+When you capture a claude.ai conversation in **mcp-store** mode, each code or
+document **artifact** Claude wrote is also saved as its own file (`parser:
+claude-ai-artifact`, tagged `artifact` + `lang:<x>`). So you can:
+
+```
+list_contexts({ tags: ["artifact"], parser: "claude-ai-artifact" })
+get_context({ slug: "...--auth-middleware", format: "code" })   // → raw code, ready to write to a file
+```
+
+That's the "one code = one file" path: ask Claude Code to pull a specific
+artifact and it gets exactly that file's contents, with no conversation or
+Markdown wrapping.
+
+### Filtering
+
+`list_contexts` and `search_contexts` share these optional filters:
+
+- `parser` — exact source, e.g. `"claude-ai"`, `"github"`, `"generic"`.
+- `tags` — case-insensitive substring match. `["lang:ts"]` matches the
+  `lang:ts` tag; `["lang:"]` matches any language tag. Multiple tags are ANDed
+  by default; pass `tagMatch: "any"` to OR them.
+- `since` / `until` — ISO date (`"2026-05-01"`) or datetime. A bare `until` date
+  covers the whole day. An unparseable bound is ignored with a warning rather
+  than failing the call.
+
+Example: *"search my claude.ai TypeScript artifacts from this month for the auth
+helper"* →
+`search_contexts({ query: "auth helper", parser: "claude-ai", tags: ["lang:ts"], since: "2026-06-01" })`.
+
+### Example
+
+In Claude Code:
+
+> *“Search my captured contexts for the OAuth migration plan and use it.”*
+
+Claude calls `search_contexts({ query: "oauth migration" })`, finds the
+`claude-ai` conversation you captured, then `get_context` to read the full plan —
+without you pasting anything.
+
+## File format
+
+The server reads the Markdown + YAML frontmatter the extension emits:
+
+```markdown
+---
+url: https://claude.ai/chat/abc-123
+title: "Auth refactor plan"
+captured_at: 2026-05-31T14:00:00.000Z
+parser: claude-ai
+tags: ["model:claude-opus-4"]
+---
+
+# Auth refactor plan
+...
+```
+
+Files without frontmatter are still served (the whole file becomes the body, and
+the title falls back to the first heading or the filename).
+
+## Troubleshooting
+
+**"I captured something but Claude Code can't see it."** This almost always
+means the server and the extension disagree on *which directory* to use. The
+server logs the exact absolute path it resolved on startup (to stderr):
+
+```
+[ccc-mcp] Watching 3 capture(s) in: /Users/you/project/.ccc-contexts
+```
+
+If you instead see `does not exist`, `is a file, not a directory`, or a count of
+`0` when you expect captures, compare that absolute path against the directory
+you linked in the extension's options page — they must resolve to the **same
+folder**. The most common cause is a relative path (`./.ccc-contexts`) resolving
+against an unexpected working directory; use an absolute path in the MCP config
+if in doubt. The `list_contexts` tool surfaces the same diagnostic to the agent.
+
+## Develop
+
+```bash
+npm install
+npm run build      # tsc -> dist/, makes bin executable
+npm test           # vitest
+npm start          # run against ./.ccc-contexts (or $CCC_CONTEXTS_DIR)
+```
+
+Logs go to **stderr** — stdout is the JSON-RPC channel and must stay clean.
+
+## License
+
+MIT