agent-discover

MCP server registry and marketplace. Discover, install, activate, and manage MCP tools on demand. Acts as a dynamic proxy -- activated servers have their tools merged into the registry's own tool list, so agents can use them without restarting.

Every MCP client today — Claude Code, Cursor, Codex CLI, Aider, Continue, plain MCP clients — requires a full agent-session restart to pick up a newly registered MCP server. The tool catalog is frozen at startup. agent-discover is the only path to register a new server and have it become discoverable in the same running session. This is the one differentiator that survives against every host, even those with their own built-in deferred-tool loaders.

Search spans the official MCP registry, npm, and PyPI in one query, so popular servers that aren't in the official index (Microsoft @playwright/mcp, @modelcontextprotocol/server-*, mcp-server-fetch, mcp-server-git, …) all show up.

Built for AI coding agents (Claude Code, Codex CLI, Gemini CLI, Aider) but works equally well with any MCP client, REST consumer, or WebSocket listener.

Light Theme	Dark Theme

Why

Static MCP configs mean every server is always running, even when unused. Adding a new server requires editing config files and restarting. There is no way to browse what is available or install new tools at runtime.

Features

• Single-call tool discovery (find_tool) — hybrid BM25 + semantic ranking returns the top match with a confidence label, compact required_args, and 4 ranked alternatives. Auto-activates the owning child server so the agent can call the proxied tool immediately on the next turn. Replaces the multi-step search → list → activate dance with one round-trip.
• Batch discovery (find_tools) — pass an array of intents to discover N tools in a single round-trip for multi-step tasks.
• Indirect invocation (proxy_call) — call a discovered tool through agent-discover without exposing it to the host catalog. Keeps the host MCP surface at exactly 5 actions regardless of how many tools the registered child servers expose — critical for very large catalogs where flooding the host with thousands of schemas would blow the model's context budget.
• Pluggable embeddings (AGENT_DISCOVER_EMBEDDING_PROVIDER) — semantic search is opt-in via none (default, BM25 only) / local (Xenova/all-MiniLM-L6-v2 via @huggingface/transformers) / openai (text-embedding-3-small). Provider failures fall back to BM25 cleanly. Mirrors agent-knowledge's pattern so the same model can be reused.
• did_you_mean recovery — when a proxied tool call fails, the proxy attaches BM25-ranked similar-tool suggestions to the error response so the agent can correct in one extra turn instead of giving up.
• Local registry -- register MCP servers in a SQLite database with name, command, args, env, tags
• Federated marketplace search -- a single query hits the official MCP registry, npm, and PyPI in parallel, dedupes by <source>:<name>, and collapses version duplicates
• PyPI integration -- curated list of well-known Python MCP servers (mcp-server-fetch, mcp-server-git, mcp-server-time, mcp-server-postgres, mcp-server-sqlite, mcp-proxy, …) plus live metadata via the PyPI JSON API; Python entries install via uvx
• npm fallback -- two parallel npm searches (keywords:mcp and <query> mcp) catch packages that didn't tag themselves (e.g. Microsoft @playwright/mcp)
• Prereqs probe -- GET /api/prereqs reports which package managers (npx, uvx, docker, uv) are available on the host; the dashboard surfaces a banner when something needed for an install is missing
• Cross-process activation -- the active flag is the source of truth in SQLite; every fresh agent-discover process hydrates its in-memory proxy from the DB on startup, so tools activated in one process show up in others
• On-demand activation -- activate/deactivate servers at runtime; their tools appear and disappear dynamically with tools/list_changed notifications
• Tool proxying -- activated server tools are namespaced as serverName__toolName and merged into the tool list
• Multi-transport -- stdio, SSE, and streamable-http transports for connecting to child servers
• Secret management -- store API keys and tokens per server, automatically injected as env vars (stdio) or HTTP headers (SSE/streamable-http) on activation; CRLF-validated to prevent header injection
• Health checks -- connect/disconnect probes for inactive servers, tool-list checks for active ones, with error count tracking
• Per-tool metrics -- call counts, error counts, and average latency recorded automatically on every proxied tool call
• Full-text search -- FTS5 search across server names, descriptions, and tags + cross-server tool index for find_tool
• Pre-download -- fire-and-forget npm cache add (npx servers) or uv tool install (uvx servers) on registration, plus a dedicated /preinstall endpoint
• Real-time dashboard -- web UI at http://localhost:3424 with Servers and Browse tabs, dark/light theme, WebSocket updates
• MCP Inspector-grade Test panel -- every active server card grows a Test drawer with seven subtabs (Tools / Info / Resources / Prompts / Events / Export / Diagnostics). Schema-driven form renderer, Pretty/Raw JSON/cURL result modes, live notification + progress streaming, localStorage presets, pop-out floating panel for side-by-side debugging, and a Test ad-hoc button that spins up a throwaway (never-registered) server with a 15-minute TTL. Covers the same surface as upstream @modelcontextprotocol/inspector without a second process or second port.
• 3 transport layers -- MCP (stdio), REST API (HTTP), WebSocket (real-time events)
• Declarative setup file -- set AGENT_DISCOVER_SETUP_FILE to a JSON file listing servers to ensure-registered on startup. Idempotent (skips existing). Supports auto_activate, env var secret refs ($VAR), and tags. Automatically also reads a .local.json variant (e.g. discover-setup.local.json) for machine-specific servers with secrets. New registry({ action: "sync" }) MCP action and POST /api/sync REST endpoint for on-demand re-read.
• Bench harness -- under bench/, comparing eager tool loading vs deferred discovery against real OpenCode + gpt-5-mini. Reproducible structural result: discover's first-turn input tokens are flat in N (~20.8k across N ∈ {10, 100, 1000, 3000}); eager's grow linearly (20.9k → 32.4k → 160.9k → context overflow at N=3000). End-to-end accuracy and multi-turn cost numbers are noisier and model-dependent — see bench/README.md for what reproduces and what doesn't.

Quick Start

Install from npm

npm install -g agent-discover

Or run directly with npx

npx agent-discover

Or clone from source

git clone https://github.com/keshrath/agent-discover.git
cd agent-discover
npm install
npm run build

Option 1: MCP server (for AI agents)

Add to your MCP client config (Claude Code, Cline, Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "agent-discover": {
      "command": "npx",
      "args": ["agent-discover"]
    }
  }
}

The dashboard auto-starts at http://localhost:3424 on the first MCP connection.

Option 2: Standalone server (for REST/WebSocket clients)

node dist/server.js --port 3424

MCP Tools (1)

A single action-based tool handles every operation via the action parameter — this keeps the prompt-overhead cost minimal regardless of how many child servers are registered.