MCP, both ways

MyDeepAgents is an MCP client. You add a server's address once; MyDeepAgents connects, asks which tools it offers, and keeps them in a list you control. A tool runs only after you approve it.

1. 01Connect

   Pick a server from the directory, paste its address, or import a config file.

2. 02Approve

   Look at each tool the server offers and allow the ones you trust.

3. 03Use

   Give the server to an agent or a workflow step, and call its tools in a run.

Open Connections → MCP servers and choose Add MCP server. Three ways in:

1. 1From the directory. Hundreds of servers, checked to be live, with their sign-in already known. Browse the directory.
2. 2By address. Paste the server's URL. MyDeepAgents probes it first and tells you whether it needs a sign-in, without storing anything.
3. 3From a config file. Paste the JSON you already use in another MCP app; see Import a config file.

What a server needs

A server belongs to one workspace. An organization can connect up to 10 servers unless its settings raise the limit.

Servers sign in in one of four ways. Whatever you enter is sealed straight away; MyDeepAgents never shows a secret again, only the names of the headers it sends.

Every tool starts as pending. It runs only once you approve it, and each approved tool is pinned: MyDeepAgents remembers exactly what the tool said it does, so a server can't quietly change a tool you trusted.

Tool states

Side effects decide when a person is asked

Each tool is marked by what it does, starting from the server's own hints, and that sets whether a run pauses for a person before calling it. You can change both for each tool.

• read→never

  Only reads. Runs without asking.

• write→policy

  Changes something. Follows your organization's approval policy.

• destructive→always

  Deletes or can't be undone. Always asks a person.

Approving in bulk

When you add a server you can approve none of its tools, only the read-only ones, or all of them. A separate rule decides what happens to new tools the server adds later.

Refresh reads the tool list again and reports what is new, changed and removed. During runs, MyDeepAgents also checks the list at most every five minutes; a tool whose description changed fails with mcp_tool_changed while the server's other tools keep working.

{
  "schema": "mda.agent/v1",
  "name": "Docs researcher",
  "slots": {
    "docs": {
      "kind": "mcp_server"
    }
  },
  "spec": {
    "model": {
      "provider": "anthropic",
      "id": "claude-sonnet-5-5"
    },
    "instructions": "Answer questions from the product documentation. Quote the page you used.",
    "tools": [
      {
        "use": "docs.search_docs"
      },
      {
        "use": "docs.fetch_page",
        "approval": "never"
      }
    ]
  }
}

• Only approved tools reach the model, each with the description that was pinned when you approved it.
• A tool's name never changes once given: it is the MCP name in snake_case. The model sees it as the slot and the tool joined by two underscores, cut to 64 characters.
• Approval comes from the agent's own tool setting first, then the server's setting for that tool, then the side-effect default. A paused run waits in Approvals.
• Turn on Share the server's instructions to add the server's own advice to the agent's instructions, quoted as the server's words and capped at 4,000 characters.

{
  "lookup": {
    "type": "action",
    "use": "docs.search_docs",
    "input": {
      "query": "${inputs.question}"
    },
    "on_error": "continue"
  }
}

What a step gets back

{
  "text": "Found 3 pages about refunds: Refund policy, Partial refunds, Refunds by card.",
  "is_error": false,
  "content": [
    {
      "type": "text",
      "text": "Found 3 pages about refunds: Refund policy, Partial refunds, Refunds by card."
    }
  ],
  "structured": null
}

On a server's page, Try calls a tool for real with arguments you type, and shows what came back. Pending and changed tools can be tried before you decide; it is a real call, never retried.

curl -X POST "https://api.mydeepagents.com/v1/mcp_servers/mcp_01jc832gmyrb97hnzx3h75bsfd/tools/search_docs/call" \
  -H "Authorization: Bearer $MDA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "arguments": {
      "query": "refunds"
    }
  }'

The call log

Every call from agents, workflows and tries is logged: the tool, the outcome, how long it took, which run made it and who. The arguments and results are never stored. The log keeps 30 days, with a daily chart of outcomes.

Limits keep one slow or chatty server from holding up a run.

Everything on the MCP pages is in the API too, under /v1/mcp_servers. Add servers from a deployment script, approve tools from a review tool, or read the call log into your own monitoring.

Add a server

curl -X POST "https://api.mydeepagents.com/v1/mcp_servers" \
  -H "Authorization: Bearer $MDA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "workspace_id": "ws_01jccj30nr8ad7mx1yc5v9zx3h",
    "name": "Linear",
    "url": "https://mcp.linear.app/mcp",
    "transport": "auto",
    "auth_type": "oauth2",
    "approve": "read_only",
    "auto_approve": "none"
  }'

Check a server before adding it

The probe connects without credentials and stores nothing. It answers whether the server is reachable, which protocol it speaks, a preview of its tools, and, if it needs a sign-in, which kind.

curl -X POST "https://api.mydeepagents.com/v1/mcp_servers/probe" \
  -H "Authorization: Bearer $MDA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://mcp.linear.app/mcp",
    "transport": "auto"
  }'

Approve tools

Send up to 500 decisions at once. Each names a tool and can set its state, its side effects and when to ask a person.

curl -X POST "https://api.mydeepagents.com/v1/mcp_servers/mcp_01jc832gmyrb97hnzx3h75bsfd/tools/decide" \
  -H "Authorization: Bearer $MDA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "decisions": [
      {
        "name": "search_issues",
        "status": "approved",
        "side_effects": "read",
        "approval": "never"
      },
      {
        "name": "delete_issue",
        "status": "blocked"
      }
    ]
  }'

{
  "mcpServers": {
    "linear": {
      "url": "https://mcp.linear.app/mcp"
    },
    "acme": {
      "type": "http",
      "url": "https://mcp.acme.com/mcp",
      "headers": {
        "Authorization": "Bearer ${input:acme_token}"
      }
    }
  }
}

Servers that run a local command can't be added and are marked so. A single Authorization: Bearer header becomes a bearer token; other headers become custom headers. mcp-remote bridges are read for the address and headers they wrap, and environment variables are filled from the server's own env.

Your organization runs its own MCP server. Switch on the agents, workflows and connection actions you want to offer, and any MCP app can call them with one of your API keys. Nothing is offered until someone switches it on.

1. 01Switch tools on

   In Developers → MCP → Your MCP server, choose Offer a tool and pick an agent, a workflow or a connection action.

2. 02Make a key

   Create an API key whose scopes include agent.run, workflow.run or connection.use, for the kinds of tools the app should call.

3. 03Connect the app

   Add the server's address and the key to the app. It lists your tools and calls them like any others.

The server answers at /v1/mcp on your API address, over Streamable HTTP. It speaks MCP versions 2025-11-25, 2025-06-18, 2025-03-26 and 2024-11-05, and offers tools only: no resources or prompts.

https://api.mydeepagents.com/v1/mcp

Each app has its own place for MCP servers. Put your key where the placeholder is; prefer an environment variable or the app's own secret prompt over pasting it into a file.

claude mcp add --transport http mydeepagents https://api.mydeepagents.com/v1/mcp \
  --header "Authorization: Bearer $MDA_API_KEY"

• A tool is listed only while it is switched on, its target can run, and both the key's scopes and its service account allow it. A key sees nothing outside its workspaces.
• A call waits up to 60 seconds. A run that takes longer, or waits for a person to approve something, answers with its run ID and where to read it, so the app or a person can follow up.
• Every call starts its own run, with the same approvals, budgets, policy and audit as a run started any other way. Runs started this way show An outside app (MCP) in Runs.
• Each key may make 60 calls a minute; more get 429 with Retry-After. An organization can offer up to 50 tools.
• Only API keys sign in, sent as Authorization: Bearer. Requests from a browser (any Origin header) are refused, so a web page can't call your tools with someone's key.

Offer tools with the API

Everything on the Your MCP server screen is in the API under /v1/mcp_exposures: offer, switch on and off, edit and remove tools. GET /v1/mcp_server tells an app the server's address and how to sign in.

curl -X POST "https://api.mydeepagents.com/v1/mcp_exposures" \
  -H "Authorization: Bearer $MDA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "kind": "agent",
    "agent_id": "agt_01jctdbahw9btk5p831xs974yq",
    "alias": "live",
    "tool_name": "support_triage",
    "description": "Answers questions about a customer'\''s open support tickets. Give it the customer'\''s name or email."
  }'