MCP server  ·  AI diagnosis  ·  Self-hosted

Give your AI the whole picture
of every project you run.

PostRequest already holds the standing answer to “what is this site, right now” — errors and their diagnoses, heartbeats, schema, DNS, sitemap and source, refreshed on a schedule. The built-in MCP server hands all of it to whichever MCP client you work in, through one read-only endpoint.

Seventeen read-only tools Every client, one endpoint Bring your own key Key never leaves your server
Errors are one facet

A 360° view of each client, kept current

Most monitoring tells you when something broke. PostRequest keeps the whole standing answer to what is this client — and refreshes it on a schedule rather than when somebody remembers to look.

That is what makes an AI useful here. It is not handed a stack trace and asked to imagine the rest — it reads a complete, current picture that was already assembled. Which is why a question spanning every site you run costs a sentence, not an afternoon.

  • Errors, their diagnoses and the whole fault history
  • PHP version, extensions, disk and TLS from the live heartbeat
  • Database schema, DNS records and the discovered sitemap
  • The retrieved source behind all of it — searchable by file and line
Model Context Protocol

Your assistant, with your whole estate in reach

Point anything that speaks MCP at your dashboard — Claude Code, Claude Desktop, Cursor, your own client — and it investigates your monitored sites directly, in the tool you already work in. No copy-pasting, no per-host access to arrange.

The questions it answers are not about errors

Faults are the obvious use, and the weakest one. The endpoint reaches every facet of every client at once, which makes the fleet-wide question — the one nobody asks today because answering it would cost a day — the ordinary case.

  • “Which clients are on an unsupported PHP version, and what would break?”
  • “Whose certificate expires this month, and did the last renewal work?”
  • “Which sites still call this deprecated helper, and where?”
  • “Did any client's schema change last week without a release?”
  • “Is this failure pattern present anywhere else in the fleet?”

Add it once: one endpoint at https://your-host/mcp, authenticated with a bearer prq_… token from Settings → API Access or over OAuth 2.1. Streamable HTTP on JSON-RPC 2.0, so it works with any MCP client — and a client-bound token scopes an assistant to a single site when that is what you want.

Why it is quick, and cheap

An afternoon's work, in one question

Ask “which of my clients need attention this week” without a tool like this and it is not one question. It is: which host, whose credentials, where does this one keep its logs, what PHP is it on, dump the schema, dig the DNS, find the release that lines up. Per client. And an assistant cannot help with any of it, because it cannot reach any of it.

PostRequest already did the collecting. The answer is read from data standing there rather than assembled while you wait — so the sweep across every site you run costs one question and a few seconds.

It is cheap on tokens for the same reason. An assistant looks for an existing diagnosis before it reads any source, so a fault explained once is not paid for twice — and source arrives by the file and by the range, never by the repository.

  • Nothing is crawled while you wait — the collecting already happened
  • A diagnosis already computed is reused, not recomputed
  • Source by file and by range — not a repository dumped into a prompt
  • One endpoint for the whole fleet — no per-host access to arrange
  • Your own key, billed by your provider, with spend broken out per feature
AI diagnosis

Every error explained before you ask

Put a key for the LLM of your choice in the dashboard — OpenAI or Anthropic — and PostRequest diagnoses captured errors on demand from the wand button, or automatically in the background as new faults arrive.

Grounded in your actual code — the stack trace, the request context, the real source around the fault — it reads like a senior engineer who already knows the codebase triaged the ticket. It is also why the rest is cheap: that written verdict is what an assistant reads first, here and over MCP alike.

  • Root-cause analysis, suggested fix and a confidence score
  • Ask follow-up questions in a chat on any error
  • API key stored encrypted at rest — never leaves your host
  • Fully optional — the monitoring works without it
What the AI can see

Look but don't touch — enforced, not promised

Seventeen tools. Every answer is anchored to what you actually run, and every one of them can only look — it can list, read, search, correlate and explain, and it can never write a file, run a command, deploy, or change a client. Point it at the whole estate and walk away: there is no approval step, no rollback plan, and nothing it did that you have to go and check.

Errors & diagnoses

Three tools · browse, search, read the verdict

List and search faults across one client or the whole fleet, and pull the existing AI diagnosis before anything is re-read.

Health & lifecycle

Two tools · heartbeat & status history

PHP version, extensions, disk and TLS from the latest heartbeat — plus how the client's status has moved over time.

Database schema

One tool · tables, columns & indexes

Inspect a client's structure to reason about data-shape questions and to spot schema drift between releases.

DNS & sitemap

Three tools · records, TLS, discovered URLs & a live fetch

Check DNS records, certificate expiry and the discovered URL map when triaging reachability or routing — and fetch a page live, because whether a site is up is answered by looking rather than by reasoning from a heartbeat.

Source code

Five tools · find, search, read by range or whole

Pinpoint the exact line behind a fault, or sweep every client for a pattern — opt-in, and only the files and ranges needed. A whole file comes back byte for byte when you want one to save, and is refused outright when it carries redacted values, because saving that copy would write the placeholder over the real setting.

What changed

Two tools · since last week, and against another client

PHP version, extensions, settings, schema and files — each entry the difference between two recorded snapshots. The answer to “it worked on Friday”, and to “it works here but not there”.

Fleet overview

One tool · every monitored client

List monitored clients so the AI can pick the right one — or ask the same question of all of them at once.

Reading source is opt-in on top of that, it stays encrypted at rest on your own server, and every suggestion cites the file and line it came from — so you can check it rather than trust it. Your key, your model, nothing handed to a third party to train on.

One engine, three ways in

The same picture, whichever way you reach it

Diagnosis, MCP and the Agents are not three products. They are three channels onto one engine — the same seventeen read-only tools, over the same collected data. What changes is who is asking, and how.

The dashboard asks for you

Nobody has to start it

Errors are diagnosed as they arrive, and the verdict is waiting when you open the page. The zero-effort channel — turn a key on and stop.

You ask, over MCP

Your client, your question

The whole estate in reach of the assistant you already work in — for the fleet-wide questions nobody asks today because answering them costs a day.

Your people ask, by email

A team on the same data

The Agents take work from clients and colleagues by email and answer it themselves — the same picture, reached by anyone who can write a message.

Take one, or all three. They share a server, a licence and a key, and the dashboard breaks the spend out per feature — so adding a channel never means adding an integration.

See what the Agents do with it →

All of it is optional, and all of it is yours. Nothing calls an LLM until you switch it on, the key never leaves your host, and every tool is read-only — so pointing an assistant at PostRequest cannot put a client at risk.

Get in touch

Want this watching your PHP sites?

Leave your email and we'll reach out to get you set up — AI diagnosis, MCP and all. No account, no obligation.

We'll send one email to confirm the address is yours — nothing reaches us until you click it. No spam, never shared.