Let your AI app use your tools, safely.
Connectors (MCP servers) let Claude and other AI apps search papers, read files or query data. pMCPg runs on your computer, holds the sign-ins, and masks personal data in what your connectors send back to the AI.
curl -fsSL https://github.com/baobab-tech/pMCPg/releases/latest/download/install.sh | shEach app keeps its own list, keys and sign-ins.
Nothing shows what the AI read or sent.
claude.ai and ChatGPT connectors live in the vendor's cloud, under your vendor account.
One question, two connectors
Follow one request from Claude, through pMCPg, to DevExplorer and your CRM and back.
VS Code and LM Studio connect to the same gateway.
search_tools call_toolYou ask Claude a question that needs two connectors.
New to connectors
How it works
- 01
Install pMCPg
One line in a terminal. It starts at login and opens its web UI.
- 02
Add a connector
Paste its URL. If it needs a sign-in, you sign in once in your browser.
- 03
Connect your AI app
One click for Claude Desktop, Claude Code, VS Code or LM Studio.
- 04
Ask your AI
It searches pMCPg for the right tool and calls it. Every call shows in the activity log.
What it does
One local process sits between your AI apps and your connectors. Every call passes through it.
One connection
Two tools, search_tools and call_tool, in front of any number of connectors.
Keys in your keychain
OAuth sign-in and token refresh. The AI never sees a key.
Personal data maskingBeta
Masks emails, phone numbers and IDs in tool results before a cloud model reads them. It doesn't touch what you type into the chat. It can miss things.
Profiles
One address per purpose. An agent sees only its profile's tools.
Approvals
Allow, ask or block each call, with one plain sentence saying why.
Tool pinning
A tool that changes after you added it is quarantined until you review the change.
Secret scanning
API keys and tokens in results or arguments are redacted. Calls carrying one are blocked.
Activity log
Every search and call, encrypted, values left out. 30 days by default.
Privacy
What leaves your machine
It depends on where your AI model and your connectors run. pMCPg talks only to your connectors and their sign-in services: no account, no telemetry, no update ping. What you type into a cloud AI app goes to its provider, as always.
Local model · local connectors
What leaves
Nothing
Suggested setup
Masking off
Masking is in beta. It catches common patterns like emails and phone numbers, and can miss names in free text. Privacy details
Teams and IT
Control for small teams today
pMCPg has no enterprise edition yet. This is what works now, and what the roadmap lists. Everything stays self-hosted: Baobab Tech runs no service and holds no data.
Works today
- Profiles with their own URLs and tokens
- Policy presets (open, balanced, careful) and custom rules
- Approvals in the web UI or the terminal
- Privacy level per profile, secret scanning, tool pinning
- Encrypted activity log, JSON-lines export
- Docker images: slim (remote connectors) and full (local ones too)
Coming · ideas on the roadmap
- Managed setup through MDM or a script
- Org policy packs: one signed file
- A private catalog of approved connectors
- A team gateway on your own server
- Audit summaries to your log system
- Keys from your secret manager
Install
One binary for macOS, Windows and Linux
The script checks the download against its SHA-256 sum, sets pMCPg to start at login, and opens the web UI. Run it again to upgrade.
curl -fsSL https://github.com/baobab-tech/pMCPg/releases/latest/download/install.sh | shWorks with Claude Desktop, Claude Code, VS Code, LM Studio, and other apps that can run an MCP command.
claude.ai, ChatGPT and Gemini in the browser can't reach pMCPg yet: they call connectors from their own cloud. It's on the roadmap.
Version 0.1 binaries are not code-signed. The one-line installer handles this. All downloads