dylr.dev How it works

How it works. One URL between your AI and your dialer.

DylrDev is a hosted MCP server. You connect your contact center platform once, paste one URL into the assistant you already use, and ask in plain English. We run the tool calls, enforce the safety tiers on our server and keep a log of every call.

Create a free account Read the setup docs

Three steps, about five minutes. No servers to run, no code to write.

  1. Connect your platform

    Create a dedicated API user or key with only the access you want your AI to have, then paste it once. We prove it works with a real read call, encrypt it with its own key and never show it again.

    Guides for all seven platforms.

  2. Add one URL to your AI

    Every connection gets its own MCP URL. Claude and ChatGPT sign in with OAuth. Claude Code, Cursor, Gemini CLI and Copilot in VS Code can use OAuth or a CLI token for that one connection.

    https://dylr.dev/mcp/your-connection

  3. Just ask

    "Build me an IVR with business hours." Reads run right away. Writes make your AI client ask you first. Risky actions wait until you approve them in the web app or arm the connection.

    See safety tiers.

What happens on every call. The same checks, in the same order, every time.

  1. Your AI sends a tool call

    For example list_campaigns or publish_flow, to the URL of one connection. A token for one connection cannot reach another.

  2. We check who is asking

    The token, the organization, your role and your access to that connection. Then the rate caps: 120 calls a minute and 5,000 a day per connection by default.

  3. We look up the tool's tier

    Read runs. Write runs, and its label tells your AI client to ask you first. Dangerous is refused unless the connection is armed or you approved that exact call.

  4. We call your platform, and only your platform

    The credential is decrypted in memory for this one request and sent only to the host you picked at setup. Any other host is refused.

  5. We write one audit row

    Time, connection, tool, tier, outcome and the AI client's name, with arguments reduced to an approved list of safe fields. Kept 90 days on the Free plan and 1 year on Unlimited, exportable as CSV.

  6. The result goes back to your AI

    We pass it through and keep nothing of it beyond a short summary such as "returned 25 records".

What your assistant sees. And what never leaves the web app.

Your AI gets

  • The list of tools for that one connection, each labeled with its tier.
  • The results of the calls it makes.
  • A link to the web app when something needs your approval.

Your AI never gets

  • Your platform credentials. They are encrypted and never returned by any API.
  • The Arm button or the Approve button. Those live behind your sign-in and a fresh password check.
  • Your audit log, your team settings or anyone else's connections.

We host MCP and nothing else

There is no DylrDev chat and no model of our own. Your assistant is the only chat window, so we never see your conversations, only the tool calls it sends.

Works with the AI you already use. Any client that speaks MCP over Streamable HTTP.

Built for teams, too. Share a connection without sharing the key.

Four roles

Owner, admin, member and viewer. Members only reach the connections they are given, and nobody sees credentials.

A second approver

Turn it on and a dangerous call has to be approved by a different admin or owner than the person whose AI asked.

Who did what

Every AI token belongs to one person, so every audit row says whose AI acted. Removing someone revokes their tokens at once.

Roles, arming and approvals in detail

Connect your first platform. Free, no card, no sales call.

Create a free account See the platforms