> ## Documentation Index
> Fetch the complete documentation index at: https://cobalt-55-abhishek.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Debugging

> Isolate MCP tool failures with MCP Inspector and MCP Logs.

When an agent doesn't see the tools you expect, or a tool call fails silently, MCP Inspector and MCP Logs cover most cases:

* **[MCP Inspector](https://github.com/modelcontextprotocol/inspector)** connects to your Refold MCP server before you wire it into an agent, lists its tools, and lets you call them with any input.
* **[MCP Logs](/v3/mcp/operate/mcp-logs)** is the audit panel in the Refold dashboard. It shows every tool call your agent made, the exact input it sent, and what came back.

Use MCP Inspector to verify the server is healthy. Use MCP Logs to trace what the agent actually did once it's running.

## MCP Inspector

\[MCP Inspector is the Model Context Protocol project's open-source debugger for any MCP server.

### Connect to your server

<Steps>
  <Step title="Start the Inspector">
    ```bash theme={null}
    npx @modelcontextprotocol/inspector
    ```

    The Inspector opens a local web UI.
  </Step>

  <Step title="Set the transport">
    In the server connection pane, select **Streamable HTTP** from the transport dropdown.
  </Step>

  <Step title="Paste your Server URL">
    Enter your Refold Server URL, copied from the server's setup page in the dashboard. The copied URL already includes your host, token, and server ID, so the `<domain>`, `<token>`, and `<server_id>` placeholders below only show the structure. Refold accepts the token two ways; use whichever is easier in the Inspector.

    <CodeGroup>
      ```bash Path token theme={null}
      https://<domain>/mcp/v1/<token>/<server_id>
      ```

      ```bash Authorization header theme={null}
      https://<domain>/mcp/v1/<server_id>
      # Add an Authorization: Bearer <token> header in the Inspector's Authentication settings
      ```
    </CodeGroup>
  </Step>

  <Step title="Connect">
    Click **Connect**. The Inspector lists the tools your server exposes and shows the JSON schema for each.
  </Step>
</Steps>

### Panels

| Panel             | What it shows                                                                                           |
| ----------------- | ------------------------------------------------------------------------------------------------------- |
| **Tools**         | Every MCP tool the server exposes with its JSON schema. Click any tool to fill in arguments and run it. |
| **Resources**     | MCP resources (empty for Refold servers).                                                               |
| **Prompts**       | MCP prompts (empty for Refold servers).                                                                 |
| **Notifications** | Server logs and raw JSON-RPC traffic.                                                                   |

### Call a tool

In the **Tools** panel, click the tool you want to test. The Inspector renders an input form from the tool's JSON schema. Fill in the fields and click **Run Tool**.

For agent mode servers, the typical test sequence is:

1. Run `RESOLVE_ACTIONS` with `integration_query: ["List Contacts Salesforce"]`
2. Copy the returned `slug`, `identifier`, `type`, and `json_schema`
3. Run `EXECUTE_ACTION`, passing `slug` as `application_slug`, `identifier` as `action_id`, the same `type`, and an `input_payload` that matches `json_schema`

See [Tools](/v3/mcp/reference/tools) for the full request and response shapes.

### The Inspector shows no tools, or not the tools you expect

If the Inspector's **Tools** panel is empty or lists tools you didn't expect, the cause is almost always the server's configuration, not the connection. Check two settings on the MCP server in the dashboard:

* **Mode.** A server runs in one mode at a time. In direct mode the Inspector lists one tool per configured action or workflow (named like `salesforce_list_contacts_action`). In agent mode it lists two meta-tools, `RESOLVE_ACTIONS` and `EXECUTE_ACTION`. If you see the meta-tools but expected per-action tools, or the reverse, the server is in the other mode. Toggle **Agent Mode** to switch.
* **Enabled actions and skills.** Only the apps and actions you added to the server appear as tools, so add any missing action on the server's **Applications** tab. The skill tools `LOAD_SKILL` and `GET_KNOWLEDGE_INDEX` (which lists the server's skills) appear only when **Retrieve Skill** is on, in either mode.

See [Server configuration](/v3/mcp/build/server-configuration) for where these toggles live and [How it works](/v3/mcp/get-started/architecture) for how each mode changes the tool list.

### Command-line usage

For scripting or smoke-testing without the UI:

```bash CLI: list tools theme={null}
npx @modelcontextprotocol/inspector --cli \
  https://<domain>/mcp/v1/<token>/<server_id> \
  --transport http \
  --method tools/list
```

```bash CLI: call a tool theme={null}
npx @modelcontextprotocol/inspector --cli \
  https://<domain>/mcp/v1/<token>/<server_id> \
  --transport http \
  --method tools/call \
  --tool-name salesforce_list_contacts_action \
  --tool-arg input_payload='{"limit":10}'
```

To send the token as a header instead of in the path, use the header-form URL and pass `--header`:

```bash CLI: token in a header theme={null}
npx @modelcontextprotocol/inspector --cli \
  https://<domain>/mcp/v1/<server_id> \
  --transport http \
  --method tools/list \
  --header "Authorization: Bearer <token>"
```

## MCP Logs

Once your agent is running, every tool call is captured in **Maintain > MCP Logs**. See [MCP Logs](/v3/mcp/operate/mcp-logs) for the full field reference, filters, and compliance details.

### A tool call returned nothing useful

<Steps>
  <Step title="Open MCP Logs and filter by linked account">
    Go to **Maintain > MCP Logs** and filter by your agent's linked account.
  </Step>

  <Step title="Find the call">
    Sort by timestamp and locate the call that should have produced the result.
  </Step>

  <Step title="Check the Status field">
    Open the detail panel:

    | Status             | What it means                                                                                                                                                                          |
    | ------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
    | `success`          | Refold completed the call. If the result still looks wrong, the problem is in how the agent interpreted the response, so compare the response payload to what the agent told the user. |
    | `validation_error` | The agent sent input that didn't match the tool's schema, or referenced an app or action that isn't on the server. The message the agent received explains what failed.                |
    | `execution_error`  | Refold couldn't reach the downstream service. The agent and the detail panel both show the same user-facing message; the full upstream error is not exposed.                           |
    | `pending`          | The call started but did not record completion.                                                                                                                                        |
  </Step>
</Steps>

<Note>
  Downstream errors from third-party apps are never surfaced verbatim. The agent and the detail panel both show the same clean, user-facing message; the full upstream error is not exposed.
</Note>

### Trace a full agent conversation

Filter **MCP Logs** by **Session ID** to see every tool call from a single agent session in chronological order. This is the fastest way to understand what an agent actually did when it called several tools in sequence, for example `GET_KNOWLEDGE_INDEX` → `LOAD_SKILL` → `RESOLVE_ACTIONS` → `EXECUTE_ACTION` × 3.

## See also

* [MCP Logs](/v3/mcp/operate/mcp-logs): full field reference, filters, and compliance details
* [Tools](/v3/mcp/reference/tools): request and response shapes for every tool
* [Connect from agent code](/v3/mcp/connect/connect-from-agent-code): wiring the Server URL into your agent
* [Authentication](/v3/mcp/connect/authentication): token model and per-user Server URLs