# Conduit

> A local-first MCP gateway that reduces agent tool-token overhead by ~90% by exposing 3 meta-tools instead of hundreds, with OS keychain secret storage and per-tool governance.

Conduit is a local-first MCP (Model Context Protocol) gateway built by South Forge AI, released under the MIT license. It solves a specific token-overhead problem: every MCP server connected to an AI agent dumps its full tool list into the agent's context on every request, and with just a few servers that can exceed 24,000 tokens before a single word is typed. Conduit sits between AI clients and downstream MCP servers, advertising only 3 meta-tools that the agent searches on demand.

## What It Is

Conduit is a native desktop application (built with Tauri + React, with a Rust gateway binary) that acts as a single local gateway for all MCP servers. Instead of each AI client loading every tool definition from every server on every request, Conduit exposes three meta-tools — `conduit_status`, `conduit_search_tools`, and `conduit_call_tool` — and routes calls to the correct downstream server at runtime. The project's own benchmark (published in `BENCHMARK.md`) reports 97% less tool-definition overhead per request and approximately 90% fewer total tokens at the same task success rate.

## How the Gateway Architecture Works

Conduit has two components working together:

- **The desktop app** — a Tauri + React UI where users manage servers, profiles, credentials, and client connections.
- **The gateway binary** (`conduit-gateway`) — launched over stdio by each AI client. It reads Conduit's local registry, connects to enabled downstream servers (stdio or remote HTTP/SSE), and routes tool calls with per-server namespacing (e.g., `stripe__list_charges`) to prevent collisions.

The registry is the shared source of truth. The gateway watches it live, so toggling a tool or adding new credentials takes effect without restarting any client. Secrets are stored in the OS keychain and injected at runtime — never written into client config files or sent to the cloud.

## Supported Clients and Integrations

Conduit detects and auto-configures a wide range of AI clients with one click, including:

- Cursor, Claude Desktop, Claude Code, Codex, VS Code, Windsurf
- Gemini CLI, Cline, Roo Code, Warp, Amazon Q, Kiro, Zed
- LM Studio, Jan, Goose, Google Antigravity

Popular MCP servers available through the built-in catalog include GitHub, Stripe, Supabase, Vercel, Linear, Notion, Sentry, and Cloudflare. Conduit can also import a client's existing server configurations.

## Governance and Observability

Conduit includes per-tool governance controls built into the desktop app:

- Toggle any individual tool on or off across all clients simultaneously
- A single switch hides every destructive tool fleet-wide
- Per-agent scoping via profiles — a coding agent literally cannot call a billing tool not in its profile
- Full audit log of every tool call with per-server latency and error rates
- A built-in tool playground to invoke any tool directly from its schema before wiring it to a client

## Update: v0.3.15

The latest release, v0.3.15 ("Tokens saved counter"), was published on June 23, 2026. The project is in active development with the core gateway, lazy discovery, OAuth/key auth with live propagation, per-tool governance, audit logging, resources and prompts proxying, and the tool playground all working end-to-end. Linux support is in beta, with the `.deb` package recommended over the AppImage for reliability. Planned future work includes team/enterprise features such as a shared/hosted gateway, RBAC/SSO, policy controls, audit export, and secret-vault integrations as a separate paid layer under an open-core model.

## Features
- Local-first MCP gateway with no cloud dependency
- ~90% fewer tokens via lazy discovery (3 meta-tools instead of hundreds)
- OS keychain secret storage — keys never in client configs
- Per-tool on/off toggles with destructive-tool global switch
- Per-agent profile scoping for server access control
- Full audit log with per-server latency and error rates
- Built-in tool playground to test tools before connecting clients
- One-click client detection and configuration for 15+ AI clients
- Hot-reload registry — credential and toggle changes propagate without restarts
- Curated server catalog plus MCP Registry search
- Full MCP proxying: tools, resources, and prompts
- Multiple accounts per service via named server instances
- No Docker, no cloud infrastructure required

## Integrations
Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Codex, Gemini CLI, Cline, Roo Code, Warp, Amazon Q, Kiro, Zed, LM Studio, Jan, Goose, Google Antigravity, GitHub MCP, Stripe MCP, Supabase MCP, Vercel MCP, Linear MCP, Notion MCP, Sentry MCP, Cloudflare MCP

## Platforms
WINDOWS, MACOS, LINUX, API, VSC_EXTENSION, CLI

## Pricing
Open Source

## Version
v0.3.15

## Links
- Website: https://conduit.southforgeai.com
- Documentation: https://github.com/tsouth89/conduit
- Repository: https://github.com/tsouth89/conduit
- EveryDev.ai: https://www.everydev.ai/tools/conduit-mcp-gateway