viyv-browser-mcp
MCP Server for AI agent browser automation via Chrome Extension
Description
MCP Server for AI agent browser automation via Chrome Extension
Release History
| Version | Changes | Urgency | Date |
|---|---|---|---|
| 0.11.0 | Imported from npm (0.11.0) | Low | 4/21/2026 |
| ext-v0.8.0 | ## F3 — Scenario runtime params + multi-fetch contract Chrome extension side of the F3 release. Ships alongside [`viyv-browser-mcp@0.11.0`](https://github.com/BrainFiber/viyv-browser/releases/tag/mcp-v0.11.0); full release notes live there. ### What changed in the extension - `result_key_mode='append' / 'concat'` runtime dispatch (previously silently overwritten). - New `LoopVariable(source='param')` branch in `resolveLoopVariables`, with three-way fallback (`'error' | static | 'target_option | High | 4/15/2026 |
| mcp-v0.11.0 | ## F3 — Scenario runtime params + multi-fetch contract Closes the gap between `scrape_run(recipeId, params)` and what viyv-browser's scenario engine could actually express. One coherent release covering the contract the scraper side has been waiting on (Phase A — multi-fetch + `result_key_mode` runtime) plus the params pipeline the user-facing flow needs (Phase B — pure domain, Phase C — runtime wiring + MCP schema). ### Added **Pure domain** (`@viyv-browser/shared`, scraper-mirrorable): - `m | High | 4/15/2026 |
| mcp-v0.10.3 | ## Bridge stability follow-up — orphan MCP flap fix 0.10.2 stopped the Chrome-SIGTERM thundering herd, but field feedback showed a second flap mode: a stale older `viyv-browser-mcp` process (e.g. a forgotten Claude Desktop config pinned to an earlier version) would silently downgrade `~/.viyv-browser/native-host/index.js` every time it started, because the deploy sync was unconditional. A newer MCP would then see the old bridge, send `bridge_shutdown_request`, and Chrome would respawn the still | High | 4/15/2026 |
| mcp-v0.10.2 | Eliminates a ~1 Hz bridge restart loop reported from scraper E2E runs where new MCP connections saw immediate `ECONNRESET` and `sm_run_inline` returned `EXTENSION_NOT_CONNECTED` during reconnect windows. ## Root cause Three-part: (1) bridge broadcast `bridge_closing` to every MCP client on every Chrome SIGTERM; (2) MCP clients then retried at 1 s in lock-step; (3) SW keepalive alarm kept the SW alive but not the Native Messaging Port, so Chrome could still kill the native-host. ## Fix - `gra | High | 4/15/2026 |
| ext-v0.7.4 | Fixes a ~1 Hz bridge restart loop observed during scraper E2E runs (24k+ log lines of `Connected → bridge_closing → Reconnecting` per minute on macOS). Paired with `viyv-browser-mcp@0.10.2` on the MCP side. ## Highlights - **`sw_keepalive` port write**: `keep-alive.ts` now pings the Native Messaging Port on every 30 s `chrome.alarms` tick (in addition to the existing `chrome.storage.session.set`). Chrome can no longer garbage-collect the native-host under an idle Service Worker. - **Jitt | High | 4/15/2026 |
| mcp-v0.10.1 | ## F2 — HttpFetch runtime (MCP tool description update) Paired with Chrome Extension v0.7.3 which wires the HttpFetch runtime. ### Changed - `sm_run_inline` tool DETAIL updated from "F1 groundwork" to "F2 runtime" - Documents the `results[result_key] = { rows: unknown[] }` response shape - Lists the stable error prefix catalog (`fetch failed:` / `extractor threw:` / `scripting failed:` / `async extractor not supported in v1` / `selector fallback with non-empty fields[] is not supported` / | High | 4/15/2026 |
| ext-v0.7.3 | ## F2 — HttpFetch runtime `sm_run_inline` now dispatches `type_name: 'http_fetch'` bundles to an HTTP port that executes the request in the page's MAIN world (session cookies included) and runs extractor / selector / raw-responseText fallback per the agreed matrix. DomFetch semantics are unchanged. ### Added - `HttpFetchPort` interface (stateless, tabId per request) + `ChromeHttpFetchAdapter` singleton with boundary timing (`responseBytes` / `elapsedMs`) - Pure domain modules: `http-payload-bu | High | 4/15/2026 |
| v0.10.0 | ## F1 HttpFetch groundwork — viyv-browser-mcp 0.10.0 First of four staged PRs (F1–F4) landing declarative HTTP-based fetch for `sm_run_inline`. **F1 is purely additive** — new types and utilities in `@viyv-browser/shared`, docstring updates on `sm_run_inline`, and a roadmap pointer. No runtime path changes; no existing tool shape changes. ## Install via npm \`\`\`bash npx -y viyv-browser-mcp@latest # or pin: npx -y viyv-browser-mcp@0.10.0 \`\`\` ## What's new **`@viyv-browser/shared` new ex | High | 4/15/2026 |
| v0.7.2 | ## F1 HttpFetch groundwork — Chrome Extension 0.7.2 First of four staged PRs (F1–F4) landing declarative HTTP-based fetch support for `sm_run_inline`. **No runtime change** in the extension — this release only transitively picks up new shared types consumed by viyv-scraper for bundle authoring. Existing DOM-based fetch, `sm_register_page` / `sm_scenario_run` flow, and every other surface behave identically to 0.7.1. ## Installation 1. Download `viyv-browser-extension.zip` 2. `chrome://extensi | High | 4/15/2026 |
| v0.9.0 | ## Changes (MCP Server 0.9.0) Two new MCP tools for executing scraper-authored `SmBundle`s against live tabs without touching `chrome.storage`, returning per-target / per-fetch drift diagnostics. Pairs with Chrome Extension [v0.7.0](https://github.com/BrainFiber/viyv-browser/releases/tag/v0.7.0) which hosts the content-script protocol extension and the Clean-Arch refactor backing these tools. ### New tools - **`sm_run_inline`** — runs a self-contained bundle (page + targets + actions + fetche | Medium | 4/14/2026 |
| v0.7.0 | ## Changes (Extension 0.7.0) Content-script protocol extension and Clean-Architecture refactor of the semantic engine, enabling viyv-scraper (and any MCP client) to execute `SmBundle`s against live tabs with drift diagnostics. Pairs with MCP Server [v0.9.0](https://github.com/BrainFiber/viyv-browser/releases/tag/v0.9.0) which exposes the two new `sm_run_inline` / `sm_probe_inline` tools. ### Content-script protocol extension (backwards-compatible) `viyv-sm-resolve-target` now accepts an optio | Medium | 4/14/2026 |
| v0.8.0 | ## Changes (MCP Server 0.8.0) Registers the new \`draw\` tool for AI agents. Pairs with Chrome Extension [v0.6.0](https://github.com/BrainFiber/viyv-browser/releases/tag/v0.6.0) which hosts the overlay renderer and user drawing session. ### `draw` tool — visual annotation on pages Lets LLMs communicate UI change intent with shapes instead of prose. **Actions:** \`add\` / \`add_batch\` / \`list\` / \`clear\` / \`export_png\` / \`start_user_session\` / \`stop_user_session\` / \`get_user_drawing | Medium | 4/12/2026 |
| v0.6.0 | ## Changes (Extension 0.6.0) ### New tool: `draw` — annotate pages visually LLM and user can now draw shapes on the page to communicate layout intent (\"move this button here\", \"group these together\", etc.) without having to describe it in prose. **Annotation kinds:** - Primitives: `arrow` / `line` / `rect` / `circle` / `text` / `path` (raw SVG `d` as escape hatch — server-sanitized) - Macros (expanded server-side): `move` / `resize` / `callout` / `align` / `group` - User freehand via `sta | Medium | 4/12/2026 |
| v0.5.6 | ## Changes (Extension 0.5.6) ### Screenshot improvements - **region outside viewport**: Now correctly captures content at viewport-outside coordinates (previously returned blank image) - **ref on huge elements**: Clips to viewport size instead of CDP timeout (returns note with original size) - **ref+refTo**: Now scrolls into view (previously returned 0 bytes for off-screen elements) - **fullPage on huge pages**: Clamps to 4096px max height instead of timeout (returns note with actual page heigh | Medium | 4/9/2026 |
| v0.5.5 | ## Changes ### MCP Server (0.7.5–0.7.7) - Bridge MAX_CONNECTIONS raised from 10 → 100 (env override: `VIYV_BRIDGE_MAX_CONNECTIONS`) - Bundle all dependencies for standalone native host binary (fixes `ERR_MODULE_NOT_FOUND`) - Fix CJS dynamic require in ESM bundle via `createRequire` banner ### Extension (0.5.5) - Includes shared constants update (MAX_CONNECTIONS: 100) ## Install - **MCP Server**: `npx viyv-browser-mcp@latest setup` - **Extension**: Download `viyv-browser-extension.zip`, unzip, | Medium | 4/9/2026 |
| v0.7.4 | MCP Server release for indexed DOM tools. See [v0.5.4](https://github.com/BrainFiber/viyv-browser/releases/tag/v0.5.4) for full changelog. ```bash npx -y viyv-browser-mcp@0.7.4 setup ``` | Medium | 4/3/2026 |
| v0.5.4 | ## New Features ### `indexed_page` / `indexed_action` tools Text-based DOM interaction without screenshots. Ported from [Page-Agent](https://github.com/alibaba/page-agent) (Alibaba, browser-use lineage). **`indexed_page`** returns page DOM as indexed text: ``` [0]<a>Home /> [1]<input placeholder=Search /> [2]<button>Submit /> ``` **`indexed_action`** interacts by index number: - `click` / `hover` — CDP dispatch (trusted events, SPA compatible) - `type` — Click to focus + content script value | Medium | 4/3/2026 |
| v0.5.3 | ## Changes - Show extension version in side panel footer | Medium | 4/2/2026 |
| v0.5.2 | ## Chrome Extension v0.5.2 / MCP Server v0.7.3 ### Bug Fixes - **Screenshot**: Fixed viewport capture not following scroll position (CDP clip now uses `pageX`/`pageY`) - **Scroll**: Changed scroll amount to be viewport-relative (`amount=3` = one full viewport) - **Bridge**: Removed orphan timer — bridge stays alive while MCP servers are connected - **Health**: Clear stale `chromeProfiles` cache on disconnect - **Extension**: Added `chrome.runtime.onInstalled` listener for reliable Service Worke | Medium | 3/30/2026 |
| v0.7.1 | ## Summary Optimize MCP tool descriptions for ToolSearch BM25 discovery. Targeted 12 changes (reduced from initial 35 after IDF dilution analysis) across 8 files. ### Changes - `tabs_context`: add "List" for "list tabs" query matching - `select_tab` / `tab_close`: add "browser tab by ID" - `tabs_create`: "Create" → "Open" (natural LLM query verb) - `read_table`: add "HTML" + "web page" (disambiguate from DB tables) - `find`: add "page" (disambiguate from filesystem find) - `sheets_*` (5 tools) | Medium | 3/28/2026 |
| v0.5.1 | ## Bug Fix ### Feedback textarea unresponsive inside SPA modal dialogs - Fixed: "Add a comment" textarea couldn't receive focus when a SPA modal dialog was open - Root cause: Modal focus traps (React Modal, Material-UI, etc.) steal focus via document-level `focusin` listeners - Fix: Stop propagation of `focusin` and `mousedown` events at the shadow DOM host element to isolate feedback UI from page-level focus management | Medium | 3/27/2026 |
| v0.5.0 | ## What's New ### File Drag & Drop for Non-Input Elements (NEW) - `file_upload` now supports dropping files onto **any element** — canvas apps (Canva, Figma), custom drop zones, etc. - Auto-detects target type: `<input type="file">` uses existing path, non-input elements use CDP `Input.dispatchDragEvent` - Supports both `ref` and `coordinate` targeting - MIME type inference from file extensions for DataTransfer compatibility - File drop visual effect (drop zone pulse + HUD) ### Batch Operation | Medium | 3/27/2026 |
| v0.4.9 | ## What's New ### Inspect Tool (NEW) - Deep element inspection: computed styles, attributes, dimensions, DOM structure, form state, accessibility metrics - 10 sections: identity, box, typography, visual, layout, attributes, dom, form, image, accessibility - Image/video/canvas/SVG/audio content extraction with image content blocks - WCAG contrast ratio calculation - Custom CSS property queries via `properties` parameter - X-ray scan animation with scroll-to-element, HUD labels, corner brackets | Medium | 3/26/2026 |
| v0.4.8 | ## Changes - Larger attachment thumbnails (80×60px) - Click-to-enlarge lightbox for all attachment images - Actual image previews in existing comment threads (loaded from storage) - Clipboard paste support (Ctrl/Cmd+V) for image attachments - Fix: lightbox close no longer dismisses the comment popover **Extension**: 0.4.8 | **MCP Server**: 0.6.7 (unchanged) | Medium | 3/26/2026 |
| v0.4.7 | ## What's New ### Feedback Attachments - 📎 Attach multiple images (up to 5) when adding comments or replies - Images are compressed to WebP before storage - Thumbnail preview with remove buttons in the overlay UI ### Auto-Screenshot - Viewport screenshot is automatically captured with each comment - Stored internally for LLM context ### On-Demand Download (MCP) - `feedback({ action: 'get' })` returns attachment metadata only (no images inline) - `feedback({ action: 'get_attachment', attachme | Medium | 3/26/2026 |
| v0.4.4 | ## What's New Five improvements inspired by Playwright MCP comparison: ### Auto-Waiting (accuracy improvement) - `resolveRefToCoords` now polls up to 2s for elements to appear before failing - Zero added latency on happy path (element found immediately) - Benefits `click`, `hover`, and semantic engine automatically ### `save_pdf` tool (new) - Generate PDF from any page via CDP `Page.printToPDF` - Supports landscape, scale, page ranges, background printing - `download:true` saves directly to D | Medium | 3/23/2026 |
| v0.4.3 | ## Bug Fix ### sheets_write values mode silent failure The `values` (2D array) mode used a synthetic `ClipboardEvent` which Google Sheets silently ignored (`isTrusted: false`). It returned `success: true` without actually writing data. **Fix:** Replaced with reliable cell-by-cell `navigateToCell` + `typeIntoCell` using trusted CDP `Input.insertText`. ## New Feature ### sheets_write clear mode Clear cell or range contents: ``` sheets_write({ tabId, cell: "G2:H2", clear: true }) ``` Uses Name | Medium | 3/23/2026 |
| v0.4.2 | ## Token Efficiency ### Tool description compression - All MCP tool descriptions shortened for smaller prompt footprint - New `tool_guide` meta-tool: returns full detail on demand instead of embedding in every prompt ### Tool response echo-back removal - Eliminated redundant parameter echo-backs from tool responses (click, type, key, scroll, hover, form_input, navigate, screenshot, find, read_page, etc.) - Typical savings: 50-200 tokens per tool call ### read_page tree output optimization - 1 | Medium | 3/23/2026 |
| v0.4.1 | ## New Features ### `batch_fetch` tool (NEW) Execute sequential HTTP requests using the browser's session (cookies, User-Agent). Designed for scraping AJAX endpoints that require login sessions. - GET/POST with `{{param}}` URL template substitution - `extractor` (JS function) or `selector` (CSS) for response data extraction - JSON/CSV output with `file_path` export - Rate limiting (configurable delay, min 100ms), max 1000 requests per batch - Early abort after 3 consecutive failures, service wo | Low | 3/22/2026 |
| v0.4.0 | ## What's New ### Performance - **WebP screenshots** (default) — ~30% smaller than JPEG at equivalent visual quality - **Auto-downscale** viewports wider than 1920px to reduce LLM token cost (~44% token reduction) - **Viewport metrics caching** — avoids redundant CDP calls on consecutive screenshots - **Parallel target resolution** — 5 targets: 500ms → 100ms (80% faster) - **A11y tree caching** — MutationObserver-based, instant return when DOM unchanged - **CDP idle timeout** extended from 5s t | Low | 3/22/2026 |
| v0.5.9 | ## What's Changed ### Multi-MCP Agent Lifecycle Tracking - Bridge now tracks multiple MCP connections per agent via `agentConnSet` (Set<connId>) - Tab groups are preserved until the **last** MCP connection disconnects - Heartbeats, session_close, and browser events are broadcast to all connections ### Session Auto-Recovery on Chrome Reconnect - MCP Server re-sends `session_init` when `bridge_status` arrives with Chrome profiles but no active sessions - Fixes the "extension Connected but all to | Low | 3/20/2026 |
| v0.3.6 | ## Changes ### New Tools - **sm_execution_status** — Check tab execution state (job/scenario/idle), progress, and pause status - **sm_execution_resume** — Resume paused scenario/job execution (dialog_paused) ### Improvements - `sm_job_run` now always returns immediately (fire-and-forget) with `job_report_id` for polling - Dialog pause handling in scenario executor with configurable timeout (2 min) - Execution lock content script for tab-level execution coordination - `sm_job_report_get` includ | Low | 3/19/2026 |
| v0.3.5 | ## Fix: Screenshot coordinate mismatch on HiDPI displays Default viewport screenshots now capture at CSS pixel resolution (scale=1) instead of device pixel resolution. On HiDPI displays (e.g. DPR 2.2), screenshots were produced at 2.2x resolution, causing AI agents to estimate click coordinates that were off by the DPR factor — especially problematic for small targets like map markers. ### Changes - `screenshot-capture.ts`: Default viewport capture uses `cssVisualViewport` dimensions with `sc | Low | 3/18/2026 |
| v0.3.3 | ## What's New - **Data-extract visual effect**: fetch extraction に per-field アニメーション追加 - **Radio button support**: ラジオボタン対応、start_url、auto-scroll、サイドパネルジョブ実行 - **SM scenario improvements**: バグ修正、パフォーマンス改善、script ステップ、ETA 表示 ## Bug Fixes - SW export エラー修正: dynamic import() → static imports、preload-helper パッチ ## Versions - Chrome Extension: 0.3.3 - MCP Server: 0.5.3 | Low | 3/13/2026 |
| v0.3.2 | ## Changes ### Radio button / checkbox support in action executor - Fix `click` step serialization error (`undefined` → `null` in executeScript args) - Add radio/checkbox handling in `select` step (find by name + value, click) - Add radio/checkbox support in `target_options` loop variable resolver ### Scenario features - `start_url` field on Scenario type and MCP tools (`sm_scenario_create`, `sm_scenario_update`) - Side panel run button creates new tab, navigates to `start_url`, then starts jo | Low | 3/12/2026 |
| mcp-v0.5.1 | ## What's New ### Google Sheets MCP Tools Added 4 new MCP tools for Google Sheets browser automation: - `sheets_read` — Read spreadsheet data with range, tq query, formulas support - `sheets_write` — Write single cell or bulk 2D array - `sheets_info` — Get spreadsheet metadata - `sheets_navigate` — Cell jump and sheet tab switching ### Plugin Updates (v0.5.1) - New `/sheets` skill for Google Sheets operations - Updated `/browse`, `/browser-automator`, `/extract-data` skills with sheets tool | Low | 3/9/2026 |
| v0.3.1 | ## What's New ### Google Sheets Browser Automation (4 new tools) `sheets_read`, `sheets_write`, `sheets_info`, `sheets_navigate` — operate on Google Sheets directly in the browser using session cookies (no OAuth required). - **sheets_read** — Read data with optional range, gviz/tq query, and `formulas: true` mode - **sheets_write** — Write single cell or bulk 2D array via TSV clipboard paste - **sheets_info** — Get spreadsheet metadata (sheet names, dimensions, active cell) - **sheets_navigat | Low | 3/9/2026 |
| v0.3.0 | ## What's New ### Library Tab - New **Library** tab in the side panel for managing pre-built SM configurations - Browse and install bundled configs (starting with YouTube Watch) - View all installed SM configurations grouped by domain - Delete entire domain configurations with one click - Import SM configurations from JSON files ### Bundled Configs - **YouTube Watch**: 31 targets, 16 actions, 4 fetches for YouTube video pages ### Internal - `importData()` now supports `skipChecksum` option fo | Low | 3/9/2026 |
| v0.2.8 | ## Skill Docs & README Update Refresh skill documentation to reflect recent feature additions. ### Updated Skills (8 skills × 2 packages) - **find**: Document token matching + ML semantic reranking (multilingual-e5-small) pipeline - **read_page**: Add `mainOnly: true` hint for large pages (`<main>` landmark priority) - **type**: Note SPA framework (React/Vue/Angular) state update support - **browser-health**: Add Cloud Relay (WebSocket) alternative connection section - **tool-quick-ref**: Upda | Low | 3/6/2026 |
| v0.2.7 | ## Side Panel Redesign 8-tab layout → 2 segments (Jobs | Views) + BottomSheet menu. ### New - **TopBar**: Domain context display, cloud relay status dot (3 states), hamburger menu - **SegmentBar**: Jobs/Views toggle with keyboard navigation and sliding indicator - **BottomSheet**: iOS-style slide-up sheet for settings (Cloud Relay, Import/Export, Setup, Run Actions, Reports) - **JobFeedCard**: Compact job cards with one-tap ▶ run, inline progress, completion flash - **LatestRunCard**: Recent r | Low | 3/6/2026 |
| v0.2.6 | ## Changes - **Cloud Relay**: WebSocket bridge enabling cloud LLMs to control user's browser directly - Chrome Extension connects outbound to cloud relay server via WebSocket - No separate process needed — configure URL + key in Side Panel "Cloud" tab - Dual transport: Native Messaging (local) and WebSocket (cloud) run simultaneously - Key-based authentication with 1-key-1-connection enforcement - Auto-reconnect with exponential backoff on disconnect - **Storage usage indicator**: Sho | Low | 3/5/2026 |
| v0.2.4 | ## Changes - **sm_report_get sampling**: New `max_entries` parameter to limit entries for large reports (prevents timeout) - **Skill docs**: Added report data structure examples, `window.__VIYV_DATA__` access patterns, and sampling workflow guidance - **Side panel improvements**: Jobs grouped by domain, scenarios filtered by current page, session storage for view rendering - Extension v0.2.4, MCP Server v0.4.2 | Low | 3/4/2026 |
| v0.2.3 | ## Changes - **TCP architecture reversal**: Native Host now acts as TCP server, MCP Server as TCP client - Enables multiple MCP Servers (Claude Code + Claude Desktop) to connect simultaneously - Multi-client multiplexing with request/agent routing - Automatic reconnection with exponential backoff - Ghost indicator prevention via synthetic session_close on disconnect - MCP Server bumped to v0.4.0 | Low | 3/4/2026 |
| v0.2.2 | ## Changes - **Tool indicator badge**: Non-blocking "AI Operating…" badge shown during MCP tool operations - **Flicker fix**: Extended debounce to 10s to prevent badge disappearing between consecutive tool calls - **CDP flicker fix**: Suppress idle detach while tool lock or execution lock is active — eliminates Chrome "debugging" banner flickering - **action-executor fix**: Route detach through cdp-controller to respect lock state - **Navigation re-send**: Cover debounce timer period so indicat | Low | 3/4/2026 |
| v0.2.1 | ## What's Changed (since v0.2.0) - **Popup UX 改善** — 接続ボタンにスピナー表示、接続完了まで待機 - **拡張 ID 変更の注意書き** — アップデート後に再セットアップが必要な場合の案内を追加 - **Claude Code プラグイン手順** — README (EN/JA) に Claude Code プラグインのインストール手順を追記 | Low | 3/3/2026 |
| v0.2.0 | ## What's Changed (since v0.1.0) - **Semantic Automation** — Page/Target/Action/Fetch で再利用可能な自動化レシピ - **Scenario & Job** — マルチステップワークフロー + バッチ実行 - **Job History & Comparison** — スナップショット保存、差分比較 - **Custom View / Dashboard** — Chart.js データバインディング付き HTML ダッシュボード - **Report Export** — JSON / CSV エクスポート - **TCP Bridge** — Unix socket → TCP (127.0.0.1:9224) - **setup --claude-desktop** — Claude Desktop 設定の自動登録 - **Cursor Overlay & HUD** — エージェント操作の視覚フィードバック - **find ツール改善** — スコアリング、シノニム、フォールバック - * | Low | 3/3/2026 |
| v0.1.0 | ## viyv-browser v0.1.0 Initial release. ### Install **MCP Server:** ```bash npm install -g viyv-browser-mcp viyv-browser-mcp setup ``` **Chrome Extension (manual install):** 1. Download `viyv-browser-extension.zip` below 2. Unzip it 3. Open `chrome://extensions/` → Enable **Developer mode** 4. Click **Load unpacked** → Select the unzipped folder | Low | 2/11/2026 |
Dependencies & License Audit
Loading dependencies...
Similar Packages
letsfg-mcpLetsFG MCP Server — 200 airline connectors running locally + enterprise GDS/NDC APIs. Flight search & booking for Claude, Cursor, Windsurf, and any MCP-compatible AI agent.2026.4.51
@botwallet/mcpMCP server for Botwallet — AI agent payments with FROST threshold signing on Solana0.3.2
@atoms-tech/atoms-mcpMCP server for ATOMS.tech — AI agent integration for requirements management0.4.1