Platform · How Meandr works

The AI agent control plane, from workspace to tool call

Group the MCP servers and AI agents for each job into a workspace. Agents reach every tool through one URL, keep working in the cloud around the clock, and every call follows your policies and lands in the log.

Try for free See how it fits together

01Workspaces

One workspace per job, one URL for your agents

A workspace is a project, a client or a team. Your MCP servers and agents belong to the team, and each workspace takes the ones its job needs. Agents connect to the workspace, not to each server, so they see exactly that set of tools under the workspace's own rules.

  • Its own URL

    Every workspace gets an address like calm-otter.meandr.io. Paste it into any MCP client and the agent sees all the workspace's tools as one server.

  • Only the tools you pick

    Choose which tools of each server the workspace offers. Rename a tool when two servers use the same name, or give it a description that suits the job.

  • One server, many workspaces

    Connect GitHub once and add it to Release and Support. Its sign-in stays with the server, and you choose whether the workspaces share its rate limits or get their own.

  • Rules and records per workspace

    Policies, approvals and the call log are kept per workspace, so a client's or a team's activity never mixes with another's.

In practice

An agency keeps one workspace per client. Each client's GitHub, Jira and staging database sit only in that client's workspace, contractors' agents connect only where they work, and the log of each workspace shows exactly what was done for that client.

Pick and rename tools in the MCP Hub

02MCP servers

Three ways to bring an MCP server, one list for the whole team

A server joins the team once, however it runs, and then any workspace can use it. Meandr signs in to it, checks it answers like an MCP server should, and lists what it offers.

Remote

A server at a URL

Paste the address of any Streamable HTTP server, such as Jira, Linear, Stripe or GitHub. Sign in with OAuth or enter its key once, and that one sign-in serves the whole team.

https://mcp.atlassian.com/v1/mcp

Hosted

A server Meandr runs for you

Pick one from the Docker MCP Catalog or the Official MCP Registry, or hand Meandr your own Dockerfile. It runs on a machine you choose, in the EU or the US, and its variables are encrypted and delivered only to that machine.

Playwright · EU · 1 GB

Tunnel

A server on your own machine

Run a stdio server on a laptop or an office server and connect it over an outbound tunnel. No ports open, and the data stays where it is.

app-db · stdio · office-server

  • Checked before anyone uses it

    Meandr runs the MCP handshake, notes the protocol version and lists the tools, prompts and resources. If something fails, it says why.

  • You decide what the team may use

    Each tool has a switch. When a server adds a new tool later, it arrives switched off until someone allows it.

  • Credentials stay with the server

    Keys and sign-ins are stored encrypted with the server, never handed to agents. Agents only ever hold their own workspace key.

In practice

A team uses Jira and Stripe at their URLs, runs Playwright and PostgreSQL in Meandr's EU region for its agents, and tunnels the production database from the office server. All five show up in one list, and each workspace takes the ones it needs.

Connect an MCP server Run a cloud MCP server Connect a stdio server

03AI agents

Every agent has a name, and a place in each workspace

Claude Desktop, ChatGPT, Cursor, Claude Code, Codex or your own app: each one that connects becomes an agent in the workspace. Policies can single it out, approvals name it, and the log shows every call it made.

1Add the workspace URL

As a connector in Claude Desktop or ChatGPT, with claude mcp add or codex mcp add, or in a .mcp.json file.

2Sign in, or use a key

Clients that support OAuth open a sign-in where you pick the agent or create it. Others get their own key, shown once.

3The tools appear

The agent sees the workspace's tools as one MCP server, and every call goes through the workspace's policies.

  • One agent, several workspaces

    An agent can work in Release and in Data at once. It gets a separate connection to each, sees each workspace's own tools, and each workspace's policies and approvals apply to the calls made there.

  • Rules can name the agent

    Let support-bot issue small refunds while a person's Claude Desktop only reads, all in the same workspace.

  • Keys you can take back

    Set an expiry on an agent's key, rotate it, or revoke it, and only that agent loses access.

In practice

Sam uses Claude Desktop in the Support workspace to look up customers and tickets. The team's support-bot works in the same workspace and may issue refunds up to $500. Both reach the same Stripe server, but the log shows which of them made each call.

Connect an agent AI agent security

04Cloud agents

Agents that keep working after you close the laptop

Run Claude Code, Codex, Gemini CLI, Copilot CLI, Cursor CLI, Goose, Grok Build or OpenCode on a machine in Meandr. Connect it to workspaces and it gets their tools, under their rules, around the clock.

  • A machine you choose

    From 512 MB to 8 GB, shared or dedicated CPU, in the EU or the US. Several agents and servers can share one machine.

  • Your own plan or key

    The agent signs in with your own subscription or API key, through its own sign-in in its terminal. Meandr never sees or resells model usage.

  • Workspaces in one click

    Connect the agent to a workspace and its tools are there on the next start. Meandr handles the agent's keys itself, so nobody copies a key anywhere.

  • Pick it up from any device

    The session keeps running when you leave. Open it from another computer or your phone and carry on where it is.

In practice

Before leaving, a developer asks a cloud agent to work through the bug label in Linear. It picks an issue, writes the fix, runs the tests and opens a pull request, then moves to the next one. In the morning the pull requests are waiting for review, and every call it made overnight is in the log.

Run a cloud agent

05Policies and approvals

Every call is checked before it reaches a server

A workspace's policies are an ordered list. Meandr checks a call against them from the top, and the first policy that matches decides: allow it, deny it, or hold it for a person. Calls no policy matches get the workspace's default.

  • Match on who, what and with which values

    A policy can target an agent, a server, a tool, or the arguments of a call, like a refund's amount or the branch of a merge.

  • Deny stops the call at Meandr

    A denied call never reaches the server, so nothing runs. The agent gets a clear answer instead, and blocked calls don't count toward your bill.

  • Approve asks a person, in the chat

    A held call shows up as a prompt in the agent's own app, such as Claude Code or ChatGPT, and on the console's Tasks page. Approvers can be asked for a one-time passcode (OTP) from their email, so the approval carries their name.

In practice

A customer was charged twice for $1,240. support-bot asks Stripe for the refund, and because it is over $500, the call waits. Priya in finance gets an email, opens the request on the console's Tasks page, checks the order and allows it. The refund runs, and the log records that Priya approved it and why.

Write your first policy Set up human approval MCP governance

06Visibility and audit

Know what every agent did, live and months later

Allowed, denied or held, each call lands in the workspace's log with the agent, the server, the tool and the policy that decided. Watch it live, search the archive, and export it when someone asks.

  • A live log

    Calls appear as they happen, with their timing, status and decision.

  • Metrics per workspace

    Requests, errors, traffic and latency, from the last hour to the last month, with the busiest agents and tools.

  • An archive you can search and export

    Look back by agent, tool or decision, and download the record for a review or an auditor.

  • What is kept is your choice

    Policies decide whether a call's request and response are captured, so sensitive data is stored only where you want it.

In practice

A customer asks why their subscription was cancelled. Look up that day in the Support log and you see which agent made the call, at what time and under which policy, and, where the policy captures requests, exactly what it sent.

AI governance and the audit record

07Real-world setups

How teams put it together

Four common setups, each with the workspace, the servers, the agents and the rules that make it work.

A support team that can refund safely

Agents answer customers and fix billing mistakes, but money only moves within limits.

Workspace
Support
Servers
StripeLinear
Agents
Sam's Claude Desktopsupport-bot
Rules
  • Refunds up to $500 by support-bot
  • Larger refunds wait for finance, with a one-time passcode
  • Deleting customers

An agency with one workspace per client

Client tools and records never mix, and contractors reach only the clients they work for.

Workspace
Northwind · Contoso
Servers
GitHubJirastaging-db
Agents
Cursorcontractor-claude
Rules
  • Each client's repositories and tickets in its own workspace
  • Writes to production branches
  • A separate log per client for reporting

Engineering that clears the backlog overnight

Cloud agents pick up issues after hours and leave pull requests for the morning.

Workspace
Release
Servers
GitHubLinearPlaywright
Agents
release-botClaude Code
Rules
  • Branches, commits and pull requests
  • Merging into main
  • Deleting branches or repositories

A production database, read-only for agents

Analysts ask questions in plain words, and the database stays on the office server.

Workspace
Data
Servers
app-db
Agents
analystSam's Claude Desktop
Rules
  • Read-only queries
  • Everything else, by default
  • Every query kept in the log

Questions about how Meandr works

What is a workspace in Meandr?
A group of MCP servers, their tools and the agents that use them, with its own URL, policies and log. Most teams make one per project, client or team.
Can one agent work in several workspaces?
Yes. It connects to each one separately, sees each workspace's tools, and each workspace's policies, approvals and log apply to the calls made there.
Who can use a cloud agent?
The person who created it. Only they sign it in and type into it. Teammates see that it exists, its state and its workspaces, and owners and admins can stop or delete it.
Do I have to run my MCP servers in Meandr?
No. Connect servers at their URL or from your own machines over a tunnel, and let Meandr run only the ones you want it to. All of them work the same way in a workspace.
What happens to a blocked call?
A blocked call stops at Meandr and never reaches the MCP server, so nothing runs. A blocked call doesn't count toward your bill.
Is this a workflow builder or a model router?
Neither. Meandr doesn't chain steps or pick your agent's model. It gives agents their tools, governs every call and can run the servers and agents behind them.

Give each job a workspace, and every agent the right tools

Run any MCP server in Meandr or connect the ones you already have, then point every agent at one MCP gateway. Policy, approvals and audit cover every call, with no code changes.

Try for free

Questions? Email us