Guide · Agent skill
How to use the Remote MCP Brain Access agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The remote-mcp skill is an instruction package for an AI agent. It walks the agent through making your personal gbrain knowledge base reachable over MCP (Model Context Protocol, a standard way for AI apps to use outside tools and data) from your other devices and from cloud-hosted agents. This guide explains what the skill does, how to install it, how to trigger it with plain requests and what to check before approving its changes.
Remote MCP Brain Accessgarrytan/gbrain · MIT Download the skill Quick start
- 1Install the skill. Download the ZIP from the page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings, Capabilities, Skills in the Claude apps.
- 2Ask in plain words. Tell your agent something like 'use my brain over MCP' so the skill starts its detection step.
- 3Check the status. Let the skill run gbrain engine status --json and gbrain mcp expose --status --json to see whether the brain is already exposed.
- 4Preview the publish plan. Review the dry-run output, which lists the changes without making them. Add --funnel only if a cloud agent needs access.
- 5Approve and publish. Say yes to the printed plan. The skill then runs the publish command, which sets up Tailscale, the HTTPS serve config and a user service, and returns an MCP URL.
- 6Register and verify each client. Create one scoped credential per client and confirm the connection works inside that client app.
What does the remote-mcp skill make the agent do?
Your brain already runs on your own computer. The skill publishes the running gbrain server so other apps can query it without moving your data. By default it uses Tailscale, a private networking service, so the endpoint is reachable only inside your tailnet and not from the public internet.
The skill also helps you register one limited-access credential for each connecting app and then checks that the connection works. The source describes the default access levels as memory-writer for ordinary memory use and memory-reader for read-only access.
- Publishes the local gbrain server with tailscale serve (HTTPS, tailnet only by default)
- Keeps the server running as a user service
- Returns an MCP URL for your clients to use
- Creates a separate scoped credential per client instead of one shared key
How do you install the remote-mcp skill?
Download the ZIP for this skill from the page, then unzip it into one of two places. Use ~/.claude/skills/ to make the skill available across your projects, or a project's .claude/skills/ folder to limit it to that project.
If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the ZIP there. The skill's package contains a SKILL.md file, which holds its instructions.
- Option A: unzip into ~/.claude/skills/
- Option B: unzip into a project's .claude/skills/ folder
- Option C: upload the ZIP in the Claude apps under Settings, Capabilities, Skills
What do you need before you start?
The skill must run on the machine that holds your brain. If you use this computer only as a thin client with no brain on it, stop and run the skill on the brain host instead. The skill installs software, changes network configuration and creates background services, so you need an always-on brain host and permission to review those system changes.
- The gbrain CLI installed on the brain host
- Tailscale, which the skill installs and signs in with your consent
- A Node or shell environment able to run gbrain serve --http as a user service
- An admin token file at ~/.gbrain/serve/admin-token, generated by the expose flow
- Optional: OAuth or machine credentials for each client, stored in a private credentials file
How do you trigger the skill with a natural request?
You do not need special syntax. Ask your agent in plain words. The skill responds to requests such as the ones below. It then starts with a detection step and checks whether your brain is already exposed before changing anything.
The skill's first step runs commands that report your engine and exposure status. If the brain is already reachable over HTTPS by another route, such as an existing ngrok or reverse proxy setup, the skill uses that endpoint rather than publishing a new one.
- 'Use my brain over MCP'
- 'Put my brain on Tailscale'
- 'Connect Claude Desktop to my brain'
- 'Reach my brain from my phone'
- 'Connect Grok Bot to my brain' (this needs a public endpoint, see the safety section)
gbrain engine status --json
gbrain mcp expose --status --jsonWhat does publishing look like?
The skill previews the plan before it changes anything. Use the dry-run form to see what would happen. Then run the real command only after you have said yes. Use the plain form for your own devices and add --funnel only when a cloud agent has to reach the brain.
Funnel makes the same name publicly reachable on the internet. The skill says to tell you before running it, and it then relies on gbrain's OAuth or bearer token authentication and scoped grants for protection. The skill's full option list, including the port setting, is in its SKILL.md file.
- Your own devices (tailnet only): gbrain mcp expose --dry-run, then gbrain mcp expose --yes
- Cloud agents (public HTTPS): gbrain mcp expose --funnel --dry-run, then gbrain mcp expose --funnel --yes
gbrain mcp expose --dry-run
gbrain mcp expose --yesWhich clients should you connect, and how?
Give each client its own credential with the least access it needs. Start with memory-writer for normal memory use, or memory-reader if the client only needs to read. The skill says not to grant broader access such as operator, full or admin to make a check pass, and it asks you before broader access.
Local agents on the same machine, such as Claude Code, Codex or opencode, use a loopback address and do not need Tailscale. Cloud agents you cannot join to your tailnet need the public Funnel route. The skill notes that a server check alone does not prove a client has loaded its tools, so confirm the connection inside the app you intend to use.
- Use one client per consumer
- Prefer memory-writer or memory-reader over broader profiles
- Check the connection from inside the actual client app
What people use it for
- Asking Claude Desktop on a laptop to search your gbrain notes while you work away from the brain host
- Letting Claude Code on another machine in your tailnet read or write memory through MCP
- Giving a cloud agent such as Grok Bot or ChatGPT a public HTTPS endpoint to query your brain, with a scoped credential
- Reaching your brain from a phone app that has joined your tailnet
Frequently asked questions
Is the remote-mcp skill free?
The skill is published under the MIT licence. The source does not list prices for the skill, and it depends on gbrain, Tailscale and your own host, so check those projects' terms for their own costs.
What do I need to install the remote-mcp skill?
You need the gbrain CLI on the machine that holds your brain, and a host that can run gbrain serve --http as a user service. Tailscale is installed and signed in by the skill, with your consent. The skill also needs permission to change network settings and create background services.
Is my brain exposed to the public internet?
Not by default. The standard publish command serves the brain only on your tailnet. Only the --funnel option makes the same name publicly reachable, and the skill says to warn you before using it.
How is this different from ngrok or a reverse proxy?
The skill uses Tailscale first and lists ngrok and cloud hosts only as alternatives. If you already serve your brain over HTTPS another way, the skill recommends using that endpoint and registering your clients against it instead of publishing a new one.
Why does the skill ask for a separate credential for each app?
Each connecting app gets a limited-access credential, so one app's access can be reviewed or removed without affecting others. The default profiles are memory-writer and memory-reader, and the skill avoids broader access unless you ask for it.
What if the check says the connection works but my app cannot see the tools?
The skill states that a server-side check does not prove a native client has loaded its tools. Re-check the connection inside the client app itself, and if the status shows pending, wait a minute and check the status again before changing anything.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you run gbrain on a machine that stays on and you want your own devices, or a cloud agent you trust, to query it through MCP with scoped access. Skip it if you have no always-on host, cannot review system changes, or want to make your data public without understanding Funnel's security model.
Download the skill