Skip to content
OnlyMCP
OnlyMCP

Every tool.One MCP.

Connect your MCP servers once. They reach every client through a single OnlyMCP connection, together with your skills, agents and memory.

npx onlymcp install --client claude-code
Get started free

Bring your own Claude Code or Codex. OnlyMCP stores no AI keys, and your service credentials never leave the server.

Claude CodeCodexAntigravitySkillsAgentsMemory
Syncing continuously
OnlyMCP speaks MCP, an open protocol. Any client that supports it can connect, including future ones. Chat clients such as Claude and ChatGPT sign in with OAuth, so there is no token to copy.
Comparison

The difference is obvious at a glance

What every machine handles on its own without OnlyMCP, OnlyMCP keeps in sync in one place.

Without OnlyMCP

  • Every MCP server is configured again in every client, on every machine.
  • Access tokens end up in config files all over your disk.
  • Skills and memory live on one machine. The second one is missing them.

With OnlyMCP

  • Connect a server once. Every client sees its tools.
  • Credentials stay encrypted on the server, never in a client.
  • A new machine needs one connection, not one per tool.
MCP connections

Connect your tools once

GitHub, Linear, Notion, your own server: whatever speaks MCP, OnlyMCP puts behind one connection, a gateway in front of all of them. Your client only ever knows about that one.

Your OnlyMCP connection

https://only-mcp.com/api/mcp

GitHub

github__

Linear

linear__

Notion

notion__

Your own server

internal__

A sample of servers that speak MCP, not built-in integrations. OnlyMCP connects to any remote MCP server that signs in with OAuth or takes a token or a header. Each connected service stays subject to its own terms, see the legal notes in the footer.

One connection, every tool

Your client is configured once. Everything you connect afterwards simply appears, without touching a config file again.

Credentials stay here

A token you store is encrypted and used by the server on your behalf. It is never sent to a client and never shown again.

You decide what is on

Turn a whole connection off, or single tools of it. In a team, permissions decide who may connect, edit and use.

Process

How it works

Three commands take you from an empty terminal to a synced device.

  1. 1

    Connect

    One command registers OnlyMCP with your client.

    npx onlymcp install --client claude-code

    OnlyMCP is now set up in Claude Code.

  2. 2

    Sign in

    Your machine shows a code, you confirm it once in the browser.

    npx onlymcp login

    This machine is signed in as a device on your account.

  3. 3

    Done

    From now on everything syncs by itself. You can check the state at any time.

    npx onlymcp status

    Prints "logged in" while this device is connected.

Synced data

What gets synced

Skills, agents and memory are the three things that stay identical across your devices.

Skills

Your skills as Markdown. The same version on every device.

Agents

Agent definitions including model and permitted tools.

Memory

Notes with paths and links, readable and writable from inside the client. Browse them in the dashboard as files or as a graph of their links.

Interface

Tools your client can call

OnlyMCP has 28 tools of its own, in four groups. Every server you connect adds its tools to the same list, so your client finds everything in one place.

Memory

8 tools

Read, write, search and move notes, and follow their links.

  • list_memory
  • memory_read
  • memory_write
  • memory_delete
  • memory_search
  • memory_tree
  • memory_move
  • memory_links

Skills

8 tools

List, fetch and edit skills, together with the files they carry.

  • list_skills
  • get_skill
  • skill_write
  • skill_delete
  • list_skill_files
  • get_skill_file
  • skill_file_write
  • skill_file_delete

Agents

4 tools

List, read, write and delete agent definitions.

  • list_agents
  • get_agent_definition
  • agent_write
  • agent_delete

Connections

8 tools

Add and manage MCP servers, switch single tools on and off, and call them.

  • list_connections
  • connection_create
  • connection_update
  • connection_delete
  • connection_tools
  • connection_tool_toggle
  • connection_tool_schema
  • connection_call

On top of these, every tool of every server you connect, named after its connection:

  • github__create_pull_request
  • github__list_issues
  • notion__notion-search
Memory

See how your notes connect

The memory view in the dashboard draws your notes as a graph: folders, projects, tags, skills and agents, joined by the links between them. This one is a small example to play with.

Launch planDecisionsOnboardingideasReleasechecklistAPI sketchResearchPricingchoiceideasWeeklyreviewMeetingInterviewnotesCodereviewerBrand voiceReadinglist

Drag a node

Connect

One endpoint, any client

Add one address to your client, then sign in or use an access token. Nothing to install, nothing that only works on one machine.

Coding agents
IDEs
claude mcp add --transport http --scope user onlymcp https://only-mcp.com/api/mcp --header "Authorization: Bearer omk_…"

In Claude, ChatGPT, Gemini, and Grok, add this address as a custom connector and sign in. Some of them offer connectors only on certain plans, the clients page says which.

https://only-mcp.com/api/mcp

Prefer a local process? npx onlymcp install still works: it is the only way to get skill files written to disk for Codex.

Setup for every client
Together

Share what the whole team needs

A team area next to your private one. You decide entry by entry what goes in.

Ownerpays per seatTeamShared in the teamSkillsAgentsMemoryConnectionsMemberPro (Team)private stays privateMemberPro (Team)private stays privateMemberPro (Team)private stays private

The owner pays per seat, and everyone else in the team gets Pro (Team) for their own account while they are in it. Requests are counted per person, so a busy member never uses up what the others have. If someone leaves, what they shared stays in the team.

See Team plan

Shared skills, agents, memory and MCP connections: the same for everyone in the team.

Five roles from owner to viewer, and single permissions you can override per person.

Your private entries stay private. Joining a team moves nothing, and whoever may delete a shared entry can return it to its owner.

Pricing

Start free

The free plan covers a small setup and costs nothing. Pay only when your setup outgrows it.

Free

CHF 0/ month

  • 3 MCP connections
  • 3 devices
  • 20 skills
  • 10 agents
  • 200 memory files
  • No teams

Pro

CHF 9/ month

plus VAT

7 days free trial

  • 25 MCP connections
  • 10 devices
  • Unlimited skills
  • Unlimited agents
  • 5000 memory files
  • No teams

Team

CHF 19per seat / month

plus VAT

  • Unlimited MCP connections
  • Unlimited devices
  • Unlimited skills
  • Unlimited agents
  • Unlimited memory
  • 5 teams
  • Any number of people per team
  • Every member gets Pro (Team)
  • Shared team connections
  • Audit log

Yearly billing is available for Pro and Team. A new account can try Pro for free.

See all plans
Security

Your data stays yours

Every row belongs to one account, every device can be revoked on its own, and a stored credential is never handed back out.

  • Every row belongs to one account, enforced in the database itself (row level security). The runtime role cannot technically bypass the rule.

  • Every connected device is listed separately and can be revoked separately.

  • Credentials for the servers you connect are encrypted at rest and never returned to any client. No AI keys with us either: you bring your own. OnlyMCP is run from Switzerland and hosted in Europe.

FAQ

Frequently asked questions

Short answers about MCP, supported clients and what OnlyMCP costs.

What is MCP?

Model Context Protocol, an open interface through which an AI client can connect additional tools and data. OnlyMCP is such a server: it provides your skills, agents and memory instead of storing them separately on every machine.

Can I connect other MCP servers?

Yes, that is the point. Add the address of any MCP server that speaks https (your own, or one of the many public ones) and OnlyMCP talks to it for you. Its tools show up in every client you already use, prefixed with the name you gave the connection. The credential stays encrypted on the server and is never handed to a client.

Which clients are supported?

Every client that speaks MCP. The documentation has checked setup instructions for Claude, ChatGPT, Gemini, Grok, Claude Code, Codex, Antigravity, opencode, GitHub Copilot (VS Code), Cursor, Windsurf, Cline, Zed, Continue, and GitLab Duo. For Claude Code, Codex, and Antigravity, npx onlymcp install writes the entry into the configuration for you.

Can I share with other people?

Yes, from the Team plan on. A team has its own area next to your private one, and you decide entry by entry what goes in. Joining a team never moves your private content.

What does it cost?

The free plan covers: 3 devices, 20 skills, 10 agents, 200 memory files. Pro (CHF 9 / month plus VAT) lifts those limits, Team (CHF 19 per seat / month plus VAT) adds 5 teams, shared connections, and an audit log. You still bring your own Claude Code or Codex: OnlyMCP never stores AI keys and never bills you for AI usage.

What happens if I lose a device?

You revoke it in the device list. That device’s access ends immediately, the others keep working.

More answers in the documentation

Ready to leave setup scripts behind?

Install the plugin, connect your first device in under two minutes.

Get Started