MCP²: Mercury Control Plane for MCP

MCP² (Mercury Control Plane) is a local-first meta-server and proxy for the Model Context Protocol (MCP). It addresses tool context bloat by enabling dynamic, progressive disclosure of tools to LLMs. Instead of flooding the model context with every available tool schema, MCP² exposes a stable, minimal surface area for tool discovery and execution.

Status

Alpha (v0.6.x) - Core functionality is implemented and tested; CLI and config details may evolve.

Install & Run

MCP² is published on npm as mcp-squared. The CLI runs on Bun (even when installed via npm), so you'll need Bun installed on your machine.

Prerequisite: Bun

Install Bun (>= 1.0.0): https://bun.sh

Run without installing (recommended)

bunx mcp-squared --help
bunx mcp-squared

Run via npm / npx

npx mcp-squared --help
npm exec --yes mcp-squared -- --help

If you see Cannot find module '@/version.js', check for a stale global install:

which mcp-squared
npm ls -g --depth=0 mcp-squared

[email protected] shipped unresolved @/... aliases in dist/index.js. Remove the old global package and re-run:

npm uninstall -g mcp-squared
hash -r
npx --yes mcp-squared@latest --help

Install globally

npm i -g mcp-squared
mcp-squared --help

# or
bun add -g mcp-squared
mcp-squared --help

Run from source

# Clone the repository
git clone https://github.com/aditzel/mcp-squared
cd mcp-squared

# Install dependencies
bun install

# Run in development mode
bun run dev

# Build for production
bun run build

# Run tests
bun test

# Evaluate tool-routing behavior (discovery-first quality check)
bun run eval:routing

# Dependency audit
bun run audit

Standalone Executable (Experimental)

Build a single-file executable with Bun:

bun run build:compile

Run the compile validation matrix (darwin/linux/windows target support, size checks, standalone smoke test, embeddings probe):

bun run build:compile:matrix

Current findings and known blockers are tracked in docs/STANDALONE-COMPILE.md.

CLI Commands (Common)

mcp-squared                 # Auto: daemon (TTY) or proxy (piped stdio)
mcp-squared --stdio         # Start MCP server (stdio mode)
mcp-squared daemon          # Start shared daemon (multi-client backend)
mcp-squared proxy           # Start stdio proxy to the shared daemon
mcp-squared config          # Launch configuration TUI
mcp-squared test [upstream] # Test upstream server connections
mcp-squared auth <upstream> # OAuth auth for SSE/HTTP upstreams
mcp-squared import          # Import MCP configs from other tools
mcp-squared migrate         # Apply one-time config migrations
mcp-squared install         # Install MCP² into other MCP clients
mcp-squared monitor         # Launch server monitor TUI
mcp-squared --help          # Full command reference

Shared Daemon Mode (Optional)

Shared daemon mode keeps a single MCP² backend alive and lets multiple stdio clients connect through lightweight proxies. This reduces duplicated upstream connections and indexing work when many tools run in parallel.

Auto mode chooses daemon when running in a TTY and proxy when stdin/stdout are piped (as MCP clients do).

mcp-squared daemon          # Start the shared backend
mcp-squared proxy           # Run a stdio proxy that connects to the daemon
mcp-squared daemon --daemon-socket=tcp://127.0.0.1:45000
mcp-squared proxy --daemon-socket=tcp://127.0.0.1:45000

Security notes for TCP daemon endpoints:

• TCP daemon bind host is restricted to loopback (localhost, 127.0.0.1, ::1).
• Optional shared-secret handshake is available with --daemon-secret=<secret> (or MCP_SQUARED_DAEMON_SECRET).

When installing into supported clients, you can register the proxy automatically:

mcp-squared install --proxy
mcp-squared install --stdio

Configuration

Config discovery order:

1. MCP_SQUARED_CONFIG environment variable
2. Project-local mcp-squared.toml or .mcp-squared/config.toml
3. User-level ~/.config/mcp-squared/config.toml (or %APPDATA%/mcp-squared/config.toml on Windows)

Minimal example:

schemaVersion = 1

[upstreams.local]
transport = "stdio"
[upstreams.local.stdio]
command = "mcp-server-local"
args = []

[upstreams.remote]
transport = "sse"
[upstreams.remote.sse]
url = "https://example.com/mcp"
auth = true

Security policies (allow/block/confirm) live under security.tools and are matched against capability:action patterns. Confirmation flows return a short-lived token that must be provided to the same capability/action call to proceed. OAuth tokens for SSE upstreams are stored under ~/.config/mcp-squared/tokens/<upstream>.json.

mcp-squared init seeds code-search routing preferences so internal retrieval/routing heuristics prioritize common code indexers by default:

[operations.findTools.preferredNamespacesByIntent]
codeSearch = ["auggie", "ctxdb"]

Tune this list (or set it to []) if your environment uses different namespaces.

Capability-first connect-time tool surfacing is configurable via inference/override settings:

[operations.dynamicToolSurface]
inference = "heuristic_with_overrides"
refresh = "on_connect"

[operations.dynamicToolSurface.capabilityOverrides]
# Optional explicit namespace -> capability pinning.
# auggie = "code_search"

For existing configs created before this default, run:

mcp-squared migrate

Use mcp-squared migrate --dry-run to preview without writing.

Migration note: capability-router mode is now the only public surface. mcp-squared migrate cleans legacy config keys and rewrites security patterns to capability/action form.

Legacy keys operations.dynamicToolSurface.mode and operations.dynamicToolSurface.naming are accepted for compatibility, ignored at runtime, and warned on load. mcp-squared migrate removes them.

Tool API (Capability Routers)

MCP² exposes one public tool per non-empty capability at connect time:

• code_search
• docs
• browser_automation
• issue_tracking
• cms_content
• design
• hosting_deploy
• time_util
• research
• general

Each capability tool uses a thin router contract:

{
  "action": "string",
  "arguments": {},
  "confirmation_token": "optional string"
}

Every capability tool supports action = "__describe_actions" for introspection. This returns capability-local action IDs, summaries, and input schemas without exposing upstream namespace/tool identifiers.

When multiple upstream instances expose the same capability action, MCP² now uses stable, instance-aware action IDs such as create_issue__github_work instead of order-dependent numeric suffixes. __describe_actions also includes baseAction, instance, and instanceTitle fields for colliding routes so clients can distinguish duplicate accounts or environments. Legacy numeric aliases such as create_issue__2 are still accepted during the transition, but the instance-aware IDs are the preferred public contract.

Recommended workflow for LLM clients:

1. Call the relevant capability (for example code_search) with action = "__describe_actions".
2. Select an action ID and call the same capability with arguments.
3. If requires_disambiguation is returned, retry with one of the returned candidate action IDs.
4. If requires_confirmation is returned, retry with confirmation_token.

Internal Retrieval Modes

operations.findTools.defaultMode supports three retrieval modes used by MCP² internals and evaluation tooling:

• fast (default): SQLite FTS5 full-text search
• semantic: Embedding similarity search (falls back to fast if embeddings are missing)
• hybrid: FTS5 + embedding rerank (falls back to fast if embeddings are missing)

Note: Semantic and hybrid modes load a local embedding model (BGE-small via Transformers.js/WASM). First load downloads ~33MB and adds ~294MB RSS. These modes are optional - fast mode (default) has no such overhead.

Embeddings are generated locally using Transformers.js (BGE-small). They are optional and can be generated programmatically via the retriever API.

Supported MCP Clients (Import/Install)

MCP² can import or install MCP server configs for: claude-code, claude-desktop, cursor, windsurf, vscode, cline, roo-code, kilo-code, gemini-cli, zed, jetbrains, factory, opencode, qwen-code, trae, antigravity, warp (import via explicit path), and codex.

Key Features

• Multi-upstream support (stdio + SSE/HTTP)
• OAuth 2.0 dynamic client registration for SSE upstreams
• Capability-first public API (one router tool per capability)
• Action-level introspection via __describe_actions
• Hybrid retrieval with optional local embeddings (FTS5 + Transformers.js)
• Selection caching with co-occurrence suggestions
• Background index refresh with change detection
• Security policies (allow/block/confirm) on capability:action patterns
• Local-first architecture (SQLite index)
• TUI interfaces for configuration and monitoring

Non-Goals

• Not a hosted SaaS platform (local-first).
• Not a replacement for MCP clients (it serves them).
• Not shipping language runtimes or SDKs at this stage.

Architecture

MCP² is built with Bun and TypeScript, leveraging:

• @modelcontextprotocol/sdk for MCP communication
• @huggingface/transformers for embeddings generation
• @opentui/core for the terminal user interface
• Zod for validation

See docs/ARCHITECTURE.md for a full architecture breakdown.

Contributing

We welcome contributions! Please see .github/CONTRIBUTING.md for details on how to get started.

Releasing

Maintainers can follow docs/RELEASING.md for the npm release checklist and docs/DEPENDENCY_MAINTENANCE.md for dependency cadence and exception handling.

License

Apache-2.0