A good MCP server gives an agent the same access the user has. It brings a human into the loop only where a mistake could be expensive, and lets the agent handle everything else on its own. Here's what that takes.
Tools the agent can pick correctly
The agent needs to choose an operation from your request and supply the right arguments. Check whether tool descriptions explain what each operation does and the arguments it accepts.
Try searching with a task such as setting up a sending domain. The results should include the matching tool and its input schema.
Bird generates its tools from the API operation catalogue. Agent descriptions explain how to use them. The dynamic MCP endpoint offers search and execute. Search returns matching tools, their schemas, effect annotations and relevant procedures. The agent can inspect a tool before calling it.
Human in the loop where a mistake could be expensive
The agent needs a human in the loop where a mistake could be expensive. Check that the review screen shows what will happen. Check whether the agent can resume from the recorded result.
Missing details need a separate question. MCP calls this elicitation: the server asks the client to get information from the human. URL elicitation opens an external interaction, such as a browser review screen.
Bird uses browser confirmation for marked operations, including number orders, eSIM orders, 10DLC registration operations and outbound voice calls. The agent prepares the request and returns a review link. You review and run the request in your browser. Accepting the elicitation opens the review. The recorded operation result lets the agent resume. The MCP setup guide explains how to connect your client.
Errors the agent can fix itself
An agent needs to know what failed, what to change and what to call next. Check a rejected field: the error should identify the problem and the action that resolves it.
For a temporary failure, the response should say whether to retry and how long to wait. After a lost response to a consequential action, the agent needs to establish what happened before trying again.
Bird's MCP error responses include a stable code, field details and a correction when one is available. The next field describes the next action. retryable and retry_after tell the agent whether and when to repeat the call. A confirmation identifier lets it retrieve the saved status and result. The API error guide describes the error fields.
Safe to act
The agent needs to distinguish reads, writes and destructive actions. Inspect the tool's effect hints and its rules for repeating a request.
An idempotency key identifies the intended action so a supported operation can return a retained response on a repeat request. Check the replay window and what the agent should do after it expires.
Bird's tools carry read-only, destructive and idempotent hints. On operations with cached response replay, the default window is three hours. Reuse the same key and request inside that window. After it expires, a retry can perform the action again. Check the operation's status before another attempt. The idempotency guide explains the operation-specific rules.
Knows the product
The agent needs product instructions while it works. Ask a setup question and check whether it can find the relevant documentation and read the complete page.
The server should also expose reusable procedures. Skills give the agent task instructions and supporting files it can load when needed.
Bird provides docs_search for ranked documentation sections and docs_read for a complete page. Its server instructions include authored procedures. Skills over MCP exposes the bird-cli and email-audit skills with readable files and file digests. The agent skills guide describes the tasks they cover.
Keeps the agent informed
The agent needs to distinguish acceptance from completion. Check whether a result includes identifiers and structured fields it can read back later.
For work that finishes later, the client should be able to receive relevant changes. A delivery notification, for example, lets the agent continue after the initial send response.
Bird returns recorded confirmation results as structured JSON. Email reads include message status, recipient status and events. Hosted MCP Events lets agents list events and subscribe to signed notifications.
Shows results
The human needs to inspect detailed results and act on them. Check whether your client renders MCP Apps, the extension for interactive interfaces inside a conversation.
A useful interface lets you inspect the underlying records or refresh their status. ChatGPT, Claude web and Desktop, and VS Code GitHub Copilot support MCP Apps rendering.
Bird's standard MCP endpoint includes DNS-verification and email-delivery widgets. You can inspect DNS records and recheck a sending domain. Email views show delivery status, recipients and an event timeline. The MCP client setup guide explains how to connect your client.
Connects in one click
The human needs a clear way to connect the agent. Check whether the client can discover the server and open sign-in.
Bird's hosted server connects through browser OAuth. The client connection instructions cover setup and sign-in.
Bird is listed in the official MCP registry and Claude's connector directory. The Bird directory instructions explain how to connect from Claude.
How to check an MCP server
Run one real task from start to finish and ask:
- Does the agent get the same access the user has, without extra setup?
- Does it pick the right tool and arguments from a plain request?
- Before anything expensive happens, does a human get to approve it?
- When a call fails, does the error say what to fix, what to call next and whether to retry?
- Can the agent retry without doing the same thing twice?
- Can it look up the product's docs and how-tos on its own?
- Can it check what happened afterwards, and hear about changes?
- Can you see the result in your client, not just read about it?