io.github.smart-mcp-proxy/mcpproxy-go MCP Proxy Server
Local-first MCP proxy that provides BM25 tool discovery, security scanning, and quarantine, with reported ~99% token savings. It routes tools for Model Context Protocol (MCP) servers and supports developer workflows using a CLI and (as indicated) a web UI.
π οΈ Key Features
Local-first MCP proxy
BM25 tool discovery
Security scanning
Quarantine
Tool routing
Token savings (~99%)
π Use Cases
Fronting an MCP server setup with a single safe endpoint
Managing and discovering tools for AI agents via BM25
Running security scanning and quarantine before tool usage
β‘ Developer Benefits
Reduced context usage via ~99% token savings
Developer tooling support (CLI) and a web UI (per README excerpt)
Focused MCP proxying and tool routing under Model Context Protocol
β οΈ Limitations
Specific implementation details beyond the provided description (e.g., supported transports, configuration options) are not included in the provided source material.
The demo above shows the embedded web UI. The MCPProxy core is a single binary for macOS, Linux, and Windows β the web UI ships inside it, with no extra service to run. On macOS, an optional menuβbar app adds oneβclick convenience (start/stop, server health, quarantine, logs).
Β Β Β Β
macOS menuβbar app Β Β Β·Β Β Activity log & audit in the macOS app
Why MCPProxy?
Scale beyond API limits β Federate hundreds of MCP servers while bypassing Cursor's 40-tool limit and OpenAI's 128-function cap.
Spend less context on tool definitions β Agents see a small set of MCPProxy discovery and execution tools (retrieve_tools, describe_tool, the call_tool_* variants and a few management tools) instead of the full upstream tool catalog, and fetch schemas on demand (in retrieve_tools and code_execution routing; direct routing exposes upstream tools individually, see Routing modes). How much context this saves depends on the number of tools, the routing mode, the client and the task; see Token savings: what is and is not measured.
Review before use β New servers are quarantined and their tools are withheld from agents until you review and approve them, and, under the default manual trust mode, most later changes to an approved tool's definition are held for review again. Known exceptions: input-schema-only edits outside scan mode, a newly added output schema, and a description added to a tool that was approved with an empty description are admitted without a hold; scan mode admits clean changes and misses what its detectors miss; auto mode does not hold changes. This reduces exposure to malicious tool definitions (Tool Poisoning Attacks); it is not a guarantee of safety. See What quarantine does and does not cover.
Pluggable security scanners β Run Snyk, Semgrep, Trivy, Cisco, and other Docker-based scanners against quarantined servers before you approve them; findings are normalized to SARIF with a composite risk score. See Security scanner plugins.
Works offline & cross-platform β A single core binary for macOS (Intel & Apple Silicon), Windows (x64 & ARM64), and Linux (x64 & ARM64), with the web UI embedded. macOS additionally ships an optional menu-bar app.
Token savings: what is and is not measured
On-demand discovery reduces how many upstream tool definitions an agent loads into context up front. The saving is not a fixed percentage: it grows with the number of connected tools and depends on the routing mode, the client and the task. MCPProxy does not publish a headline token-reduction or accuracy-improvement figure, and it makes no latency claim.
What the repository does measure, as a first-party, deterministic estimate (see bench/README.md for the method and to reproduce it with go run ./bench/cmd/bench):
Corpus: the frozen 45-tool reference corpus; counting tool name and description only (input schemas excluded), cl100k_base.
Metric: size of the initial tool menu in tokens, including the proxy's own discovery, call and management tools. Not total workflow tokens, not latency, and not task success.
Result: on that 45-tool corpus the proxy's initial menu is larger than listing every upstream tool (2625 tokens in retrieve_tools mode and 1874 in code_execution mode, against 1730 for the baseline), because the proxy's own tools have a fixed cost. A saving appears only above a break-even number of upstream tools, so quote the tool count with any percentage.
Quick Start
1. Install
macOS (Recommended - DMG Installer):
Download the latest DMG installer for your architecture:
Apple Silicon (M1/M2):Download DMG β mcpproxy-*-darwin-arm64.dmg
yay -S mcpproxy-bin
# or
git clone https://aur.archlinux.org/mcpproxy-bin.git && cd mcpproxy-bin && makepkg -si
The apt and dnf packages ship a hardened systemd unit and start the service automatically. Repository signing key fingerprint: 3B6F A1AD 5D53 59DA 51F1 8DDC E1B5 9B9B A1CB 8A3B.
For one-off .deb / .rpm downloads (air-gapped installs), grab them from the latest release.
go install github.com/smart-mcp-proxy/mcpproxy-go/cmd/mcpproxy@latest
2. Run
bash
mcpproxy serve # starts HTTP server on :8080 and shows tray
3. Add your first server and approve it
New servers are held in quarantine until you review them, so adding a server is a two-step journey: add, then review and approve. The full walkthrough, with a real remote server (Context7), is the published Quick Start. In short (Context7 is a public, read-only documentation server; substitute any server you trust):
bash
# 1. add (held in quarantine until approved)
mcpproxy upstream add context7 https://mcp.context7.com/mcp
# 2. review: capture the tool definitions, run the offline scan, read the verdict
mcpproxy review fetch context7
mcpproxy security rescan context7
mcpproxy review show context7
# 3. approve (with no --tools/--all, only read-only tools with a clean scan are allowed)
mcpproxy review approve context7 --yes# 4. one harmless read, then see it in the activity log
mcpproxy call tool-read -t context7:resolve-library-id -j '{"libraryName":"vue-router","query":"routing"}'
mcpproxy activity list --limit 3
A clean scan is not proof that a server is safe; it only means the offline checks found nothing in the tool definitions. Review what you approve.
You can also edit ~/.mcpproxy/mcp_config.json by hand; a first-seen entry is held for review in the same way:
Once connected, your agent sees a handful of built-in MCPProxy tools instead of hundreds of upstream schemas. A typical session has three beats β discover, call, audit β plus an optional preflight gate for unattended automations.
1. Discover β spend one query, not your context window
The agent asks for what it needs in plain keywords via retrieve_tools:
json
{"query":"create github issue","limit":5}
MCPProxy runs a BM25 search across every connected server and returns only the top-ranked matches β each with a call_with hint recommending the right call variant for its annotations:
This is where the token savings come from: the schemas of the hundreds of tools the agent didn't need never enter its context. The agent loads full schemas on demand with describe_tool (batch up to 5 ids) only for the tools it's about to use.
2. Call β with declared intent
The agent executes the tool through the variant matching its intent (call_tool_read, call_tool_write, or call_tool_destructive), addressing it as server:tool:
json
{"name":"github:create_issue","args":{"repo":"acme/api","title":"Bug report"},"intent_reason":"Filing bug per user request"}
MCPProxy validates the intent against the tool's annotations (a "read" call can't reach a destructive tool) and checks quarantine and approval state before the call is forwarded. Sensitive-data detection flags potential secrets in arguments and responses in the activity log; it does not block the call by default. Optional output sanitisation can redact detected secrets or block critical detections in supported text responses before they are forwarded. It does not cover tool arguments, non-text content, direct routing mode, or upstream error responses.
3. Audit β every call is on the record
Calls are recorded in the local Activity Log with a request ID, so you can inspect what an agent did. The log is bounded by configurable retention (age, record count and size) and a per-record response size cap, so it is an inspection and debugging aid, not a compliance guarantee:
bash
mcpproxy activity list # everything, newest first
mcpproxy activity list --request-id <id> # one workflow, correlated
Gate automations before they burn tokens
For recurring headless jobs (cron, CI, n8n), don't let the agent discover a missing tool the expensive way. One preflight command checks that every required tool is ready β without contacting any upstream server β and reports exactly why when it isn't (server quarantined, tool changed since approval, OAuth expired, typo'd id):
bash
mcpproxy tools preflight gh-ops:sync_issues slack:post_message --wait 10s
case $? in
0) run-agent-session ;; # all ready β go
10) exit 75 ;; # transient (server starting) β let the next cron tick retry
11) page-operator ;; # blocked β someone must approve / enable / log in
12) fail-pipeline ;; # unknown tool id β the automation itself is misconfiguredesac
See Required-Tools Preflight for the full reason taxonomy, REST endpoint, and GitHub Actions / n8n recipes.
π Optional HTTPS Setup
MCPProxy works with HTTP by default for easy setup. HTTPS is optional and primarily useful for production environments or when stricter security is required.
π‘ Note: Most users can stick with HTTP (the default) as it works perfectly with all supported clients including Claude Desktop, Cursor, and VS Code.
Quick HTTPS Setup
1. Enable HTTPS (choose one method):
bash
# Method 1: Environment variableexport MCPPROXY_TLS_ENABLED=true
mcpproxy serve
# Method 2: Config file# Edit ~/.mcpproxy/mcp_config.json and set "tls.enabled": true
2. Trust the certificate (one-time setup):
bash
mcpproxy trust-cert
3. Use HTTPS URLs:
MCP endpoint: https://localhost:8080/mcp
Web UI: https://localhost:8080/ui/
Claude Desktop Integration
For Claude Desktop, add this to your claude_desktop_config.json: