Model Context Protocol guide

Use an MCP server for API monitoring.

Connect Claude Code or OpenAI Codex to customer-aware API analytics, then investigate incidents, unhealthy integrations and release regressions without leaving your coding agent.

Coding agents understand source code. They do not automatically know which customer integrations are failing in production, whether a new release changed API reliability, or which accounts remain on an old API version.

An API monitoring MCP server closes that gap through explicit tools. The agent asks for a bounded analysis, the monitoring system computes it from existing telemetry, and the agent explains the evidence. This is materially different from giving a model direct database access or sending raw logs into a prompt.

01 · The role of MCP

Connect an AI client to operations, not to arbitrary data.

Model Context Protocol is an open standard for connecting AI applications to external tools and data sources. For API monitoring, the useful boundary is a small set of named operations rather than a general query interface.

1

Your API

Framework SDKs emit normalized route, reliability, latency, release and customer context.

2

Apirelio analytics

Deterministic services calculate customer health, incident impact and period comparisons.

3

MCP tools

Eight explicit read-only tools return bounded structured results for one project.

4

Your coding agent

Claude Code or Codex turns the returned metrics into a concise explanation.

Apirelio does not call an LLM.

The MCP server exposes deterministic analytics. Your existing AI client decides when to call a tool and how to present its result, so Apirelio requires no OpenAI or Anthropic API key.

02 · Runtime context

Source code explains what can happen. Production analytics show who is affected.

A coding agent can inspect handlers, tests and schemas. It still needs trusted operational evidence to prioritize a customer-facing incident.

Repository context

What the agent already sees

  • Route and controller implementation
  • Error handling and tests
  • Recent code changes
  • Configuration and dependencies
Apirelio MCP context

What production adds

  • Affected customers and integrations
  • Error rate and P95 latency changes
  • Release and API-version impact
  • Stalled onboarding and activity drops

Together, those contexts support a better investigation. The agent can identify the affected account and endpoint through Apirelio, inspect the relevant implementation in the repository, and propose a focused next step. The MCP result is evidence for the workflow, not a replacement for logs, traces or a safe reproduction.

03 · API monitoring workflows

Ask operational questions with customer context.

The strongest use cases combine a concrete production question with bounded dimensions and an explicit time period.

01

Prioritize customer risk

Find critical and warning accounts, then explain the activity, reliability, performance and adoption signals behind each health score.

02

Investigate an API incident

Analyze one stable error code by occurrence trend, affected endpoints, customers, releases and API versions without returning raw events.

03

Compare a release

Measure success-rate and P95-latency changes against the preceding equal-duration baseline and identify impacted customers.

04

Explain what changed

Compare traffic, error rate, latency, endpoints, customers and new errors with the immediately preceding period.

05

Track integration journeys

Review first request, first success, production adoption, time-to-value and customers stalled before regular usage.

06

Resolve customer identity

Search by an account name, retrieve its stable external ID and use that identifier in customer-specific analytics tools.

04 · Example session

From a prompt to customer evidence.

In this anonymized example, the coding agent resolves a customer, calls the customer-health tool and explains the metrics returned by Apirelio.

Coding agent calling Apirelio MCP tools to investigate an unhealthy customer integration

Illustrative sample data. The animation alternates between a customer-health investigation and a release-impact comparison.

05 · Security boundary

Read-only by design and isolated by project.

An AI integration should not turn observability into an unbounded data-access layer. Apirelio applies the same tenant boundary before any tool can run.

Dedicated credentials

MCP connections use separate high-entropy bearer tokens. Ingestion keys, dashboard sessions and general Sanctum tokens are not accepted.

One project per token

The route project, token project and analytics context must match. Cross-project requests return the same generic authentication failure.

Read-only scope

All tools carry only analytics:read. There is no generic SQL, HTTP, URL-fetch or write operation.

Bounded output

Periods, list lengths, request bodies, result size, execution time and per-connection request rate are capped.

No raw events

Tools expose aggregate operational evidence. Incident investigation deliberately removes recent raw event records.

Minimal audit

Audit rows record the project, connection, tool, status, duration and result size, never arguments, results or credentials.

The connected AI provider will process the analytics returned to its client. Review that provider's privacy and retention terms before connecting a production project. Apirelio's own telemetry model keeps request bodies, response bodies, credentials, cookies, query values, email addresses and client IP addresses out.

06 · Connect your client

Use the project URL and a one-time connection token.

Create a connection on the project's AI & MCP page. Copy the complete Server URL and token immediately. Apirelio stores only the token hash.

Claude Code

Export APIRELIO_MCP_TOKEN, place this configuration in .mcp.json, then verify it with claude mcp list or /mcp.

.mcp.jsonjson
{
  "mcpServers": {
    "apirelio": {
      "type": "http",
      "url": "https://apirelio.com/mcp/projects/YOUR_PROJECT_ID",
      "headers": {
        "Authorization": "Bearer ${APIRELIO_MCP_TOKEN}"
      }
    }
  }
}

OpenAI Codex

Export the same environment variable, add this table to ~/.codex/config.toml, then run codex mcp list or /mcp.

~/.codex/config.tomltoml
[mcp_servers.apirelio]
url = "https://apirelio.com/mcp/projects/YOUR_PROJECT_ID"
bearer_token_env_var = "APIRELIO_MCP_TOKEN"
startup_timeout_sec = 30
tool_timeout_sec = 60
enabled = true
Replace YOUR_PROJECT_ID.

It is a placeholder. If the Apirelio project page shows a URL ending in /projects/17, use 17. Copying the complete Server URL avoids configuration mistakes.

Using several projects

Create one connection per project and register each under a descriptive client-side name, such as apirelio-orders and apirelio-billing. The coding agent can use both servers in one session, while each credential remains unable to cross its own project boundary.

07 · Prompts to try

Start with a decision, account or release.

Good prompts state the operational question and let the agent select the narrowest available tools.

“Which production customers need attention this week, and what evidence explains the ranking?”
“Find Acme, then explain its integration health compared with the previous seven days.”
“Investigate PAYMENT_TIMEOUT. Which endpoints, customers and releases are most affected?”
“Did release 2026.08.4 create a regression in success rate or P95 latency?”
“Summarize what changed in production compared with the preceding period.”
“Which integrations are stalled before regular production usage, and what should happen next?”

Continue exploring

Connect the MCP workflow to your monitoring, privacy and framework setup.

Questions

Frequently asked questions.

What is an MCP server for API monitoring?+

It is a Model Context Protocol server that gives a compatible AI client access to bounded API monitoring tools. Instead of exposing arbitrary SQL or raw telemetry, Apirelio exposes explicit read-only operations for customer health, incidents, releases, changes and integration journeys.

Can Claude Code and OpenAI Codex use the same MCP server?+

Yes. Both support remote Streamable HTTP MCP servers. Each client uses the project URL and sends its Apirelio connection token as a bearer credential.

Does Apirelio send API payloads to the coding agent?+

No. Apirelio does not capture request or response bodies. MCP tools return bounded aggregate analytics such as stable customer IDs, route templates, rates, counts, releases and latency metrics. The connected AI provider then processes the specific tool results requested by the user.

Can one MCP token access multiple projects?+

No. Every connection belongs to one project and carries only the analytics:read scope. To use several projects, create one connection per project and register each endpoint under a descriptive name in the AI client.

Does the MCP integration require an OpenAI or Anthropic API key in Apirelio?+

No. Apirelio provides deterministic analytics tools and never calls a language model. The user connects Apirelio to an AI client they already use, such as Claude Code or OpenAI Codex.

Bring customer context into your agent

Investigate the API your customers actually use.

Create a project, connect your framework and give Claude Code or Codex secure read-only access to the resulting analytics.