Repository navigation
Releases
v0.6.2
Bug fix
Claude Desktop integration was silently broken. mcp-manager wrote mcpServers entries in the {"type": "http", "url": "..."} form for every tool, but Claude Desktop rejects that schema and logs Skipped invalid MCP server config entries in ~/Library/Logs/Claude/main.log. None of the proxied servers ever loaded in Claude Desktop.
Now: for the claude-desktop tool, entries are written as a mcp-remote stdio wrapper:
{
"command": "npx",
"args": ["-y", "mcp-remote", "http://localhost:PORT/mcp/UUID?client=claude-desktop", "--allow-http"]
}Other tools that share the mcpServers format (Claude Code, Cursor, Windsurf) keep receiving the native http schema they already support. The proxy-detection logic also recognizes the stdio-wrapper form so toggling the Claude Desktop integration off/on doesn't reimport our own entries as foreign servers.
Install
After installing, strip macOS quarantine:
xattr -dr com.apple.quarantine /Applications/Agent\ Hub.appv0.6.1
Bug fix
Memory MCP returned -32602 after sleep/wake. v0.6.0's eventsource-client correctly auto-reconnects when the upstream SSE socket dies, but the upstream MCP server (e.g. Redis Agent Memory) issues a fresh session per connection. We were updating the POST URL but never re-running the MCP initialize handshake on the new session, so every subsequent tools/call got rejected with "Received request before initialization was complete."
Now: the legacy-SSE reader detects each reconnect, waits for the new endpoint event, and re-runs initialize + notifications/initialized on the fresh session. Tool calls work transparently after sleep/wake or any upstream restart.
Install
After installing, strip macOS quarantine:
xattr -dr com.apple.quarantine /Applications/Agent\ Hub.appv0.6.0
Bug fixes
- Codex integration: enable/disable was broken on toml crate 1.0+ — switched to
toml::from_strso configs parse correctly again. - Memory MCP auto-reconnect: long-lived legacy-SSE connections used to become unreachable after macOS sleep/wake (or any silent socket death), requiring a manual disable/re-enable. Replaced the hand-rolled SSE reader with
eventsource-client— it auto-reconnects with exponential backoff and uses a 60s read timeout to detect half-dead sockets.
Since v0.5.0
- Directory-based profile system (per-project config)
- Cloudflare tunnel integration (enable/disable + status in UI)
- Remote Redis support for the Memory server
- Cursor MCP reconnect hint after restart
Install
Apple Silicon, unsigned build. After installing, strip macOS quarantine:
xattr -dr com.apple.quarantine /Applications/Agent\ Hub.appv0.5.0
What's Changed
- Fix: Claude CLI path resolution in DMG builds —
resolve_claude_binary()is now shared between plugins and integrations, so the Claude Code integration works correctly when launched from Finder/Dock (not just terminal).
Downloads
- Agent Hub_0.5.0_aarch64.dmg — macOS Apple Silicon
v0.4.0 — Agent Hub
Agent Hub v0.4.0
Rebrand: MCP Manager → Agent Hub
All internal references now match the public brand name. This is a cosmetic rename — no functionality changes.
What changed
- App name, window title, tray menu, and OAuth pages now say "Agent Hub"
- Bundle identifier changed to
com.agenthub.app(store auto-migrates from old ID) - Docker containers renamed to
agent-hub-*(export memories before upgrading, then re-import) - Installed skills migrated from
mcp-manager/→agent-hub/source on first launch
Upgrade notes
- Memory users: Export your memories before upgrading. After launching the new version, re-import into the new
agent-hub-*containers. - Config store migrates automatically from
com.mcpmanager.app→com.agenthub.app.
v0.3.0
What's New
Plugins & Skills Marketplace
- Browse and install MCP skills from the marketplace
- Claude Code plugins: browse, install, enable/disable with one click
- Local skill discovery: detects skills already on disk (
~/.claude/skills/) - Cross-references local skills in marketplace search results
- Markdown rendering for skill detail views
- Green checkmark badges on installed marketplace cards
Managed Skills
- Memory and Tool Discovery features now auto-install their companion skills
- "Managed" badge in UI for auto-managed skills
- Managed skills protected from accidental uninstall
- Skills auto-enable/disable when their parent feature toggles
Skill Integrations
- Per-skill integration detection (which AI tools support the skill)
- Enable/disable skill integrations per AI tool
- Import-on-enable: toggling a skill integration writes it to the target tool's config
Memory Browser Improvements
- Consolidated memories store with search logic moved from composable
- Import/export/format memories via Data Management view
- Better error handling for memory server connectivity
- OpenAI as default embedding provider, with duplicate migration warnings
- Infinite scroll and stable ordering fixes
Tool Discovery Mode
- Single-endpoint LLM tool access for simplified integrations
- Discovery skill auto-installed as managed skill
Navigation & UI Overhaul
- Router-based navigation: URL is source of truth for selection state
- Sidebar uses
router-linkwithactive-class— no more manual highlighting - Collapsible sidebar sections for Servers, Skills, Plugins, and Memory
- Marketplace merged into Add Server page
- Larger default window (1200x800)
Reliability
- OAuth tokens persisted across restarts
- Auto-detect HTTP proxy wrappers for servers
- Fix race condition on restart for tools and logs
- Unified managed MCP configs across all AI tools
- Per-server proxy endpoints with dynamic integration config sync
- Streamable HTTP proxy compliance for Cursor compatibility
Code Quality
- Vue Router and Pinia conventions documented in CLAUDE.md
- Stores cleaned up: removed selection state, let router own navigation
- Frontend types reorganized by domain
- Rust backend: split state module, fix mutex-across-IO, DRY persistence
v0.2.0
What's New
System Tray
- macOS menu bar tray icon with server status indicators
- Quick connect/disconnect actions from the tray
- Click server name to navigate directly to its detail view
AI Tool Integrations
- Auto-detect installed AI tools (Claude Desktop, Cursor, Windsurf)
- One-click enable/disable to import existing MCP server configs
- Proxy endpoint auto-updates in tool configs on startup
Managed Memory Server
- One-click toggle in Settings to enable shared agent memory
- Spins up Redis Stack and Ollama Docker containers automatically
- Uses local
nomic-embed-textmodel for embeddings — no API keys needed - Memory tools (search, store, etc.) proxied to all connected AI tools
- Managed servers protected from accidental deletion
Skills Browser
- Browse prompt/skill templates exposed by connected MCP servers
- Collapsible Skills section in sidebar
Auto-Connect & Navigation
- Servers auto-connect on app launch
- Collapsible sidebar sections (Servers, Skills)
Server Logging
- MCP server stderr output now flows to frontend LogViewer
- Smart log level detection for Python servers (INFO/WARN/ERROR)
- Proxy tool call traffic logged at info level
- Request timeout bumped to 60s for slow-starting servers
Code Quality
- Extracted shared
ServerForm.vuecomponent for add/edit views - Extracted
useServerStatus.tscomposable - Reorganized types into domain files (oauth.ts, proxy.ts, integration.ts)
- Removed dead code, replaced
.unwrap()with.expect() - Updated CLAUDE.md with architecture conventions
MCP Manager v0.1.0
MCP Manager v0.1.0
A native macOS app for managing MCP (Model Context Protocol) servers.
Features
- Add, edit, delete, and toggle MCP servers
- stdio transport (spawn local processes)
- HTTP transport (streamable HTTP + legacy SSE)
- OAuth 2.1 authorization with PKCE for authenticated servers
- Tool browser and log viewer per server
- MCP proxy server for aggregating tools across servers
Download
- MCP Manager_0.1.0_aarch64.dmg — macOS Apple Silicon (arm64)
Requirements
- macOS 13+ (Ventura or later)
- Apple Silicon (M1/M2/M3/M4)