The instrument that explains what your agents did.

Mirador watches every AI system you run and the agents inside each one, whatever built the agents, scores health and compliance, tracks the cost, and explains each decision an agent made.

Runs as containers in your own cloud, reads what your agents already record, and sends nothing back to us.

Open the consoleTalk to the founder

Overview

This weekThis monthThis quarter
Updated 12 s agoExport PDF

2 critical1 warning3 healthy6 systems, 18 agents

Success rate
98.4%▼ 0.7%
This week98.4%
Last week99.1%
Runs
57,580▲ 3.2%
This week57,580
Last week55,790
Spend
£9,583▲ 9%
Month to date£9,583
Last month£8,760
Compliance score
85▲ 3
This month85
Last month82

Estate

Bigger stars hold more agents
AI SDR agents8,400 runsSupply chain agents17,940 runsCustomer support agents24,600 runsFinance ops agents5,200 runsHiring agents260 runsMarketing agents1,180 runs

Priority actions

3 of 7
  • Supply chain agentsReliability

    Turn retry on for the OpenAI step

    1 in 25 runs fails there. 3 of 4 failed calls succeeded on retry.

    1 in 25 runs recovered

  • AI SDR agentsEU AI Act 14

    Add an approval step before outreach sends

    Lead details reach OpenAI with no human check.

    Closes 1 finding

  • Customer support agentsCost

    Move reply drafting to Gemini 3.8 Flash

    38 of 40 sampled replies matched GPT-6 Astra. 13 times cheaper per token.

    Save £1,140 a month, £13,700 a year

Critical4 agents · n8n · Owner Priya Raman · 3rd worst of 6 systems this week, and the only one getting worse

Your workflow

Open in n8n
Your cloud · Azure uk-south · account 4471Outside your cloudPrompt and responseLeaves your cloud1 open findingsupplier dbmailboxSupplier triagePO matchingDelivery exceptionsFreight deskledgerDeepSeek V4.1self-hostedGPT-6 AstraOpenAI
2Match to invoiceAI agentHealthy1.1 sAgent reasoning

Saw

12 PO lines. Invoice INV-20931 from the ledger. The instruction: match by SKU, then by description, never guess on a supplier name.

Reasoned

Lines 1 to 11 match INV-20931 by SKU. Line 12 has SKU KC-8871, but the invoice names the supplier as Kestrel Comp. Ltd and the purchase order says Kestrel Components. I am told never to guess on a supplier name, so I will hold line 12 and use the reconcile tool to compare the two names.

Did

Matched 11 of 12 lines. Marked line 12 for reconciliation.

Step 3, because the agent chose the reconcile tool for line 12

Estate6 systems · 18 agents

Bigger stars hold more agents. Tap a star to open the system.

Critical4 agents · n8n · Owner Priya Raman · 3rd worst of 6 systems this week, and the only one getting worse

2Match to invoiceAI agentHealthy1.1 sAgent reasoning

Saw

12 PO lines. Invoice INV-20931 from the ledger. The instruction: match by SKU, then by description, never guess on a supplier name.

Reasoned

Lines 1 to 11 match INV-20931 by SKU. Line 12 has SKU KC-8871, but the invoice names the supplier as Kestrel Comp. Ltd and the purchase order says Kestrel Components. I am told never to guess on a supplier name, so I will hold line 12 and use the reconcile tool to compare the two names.

Did

Matched 11 of 12 lines. Marked line 12 for reconciliation.

Step 3, because the agent chose the reconcile tool for line 12

A black hole, labelled your AI agents. A small star, labelled your data and PII, is pulled into a stream that wraps the hole, winds inward and fills a disc of hot gas, brighter on the side that turns towards you, while the view comes down from above the stream to the plane of the disc. The disc turns for a moment. Then the whole swallow plays backwards: the disc unwinds into the stream, the stream runs back out into the star, and the star stands whole where the loop began, before the hole takes the star again.

Your agents are black holes.

Data goes in and decisions come out. Nobody can see what happened in between, and when a regulator asks why an agent did something, “it ran” is not an answer.

A black hole cannot be seen. Its edge can.

Mirador sits at the edge, records every prompt, tool call and reply on the way in, and replays each run step by step: what the agent saw, what it reasoned, what it did, and why.

Every AI system you run on one page, with health, cost and compliance worked out from the same record.

Four things a leader needs to know about an AI system, on one page.

  1. Every system gets a status. The status comes from success rate, duration, cost per run, tool errors and the quality of what the agents produced, against the system's own history. One colour, one word, then the numbers.

Supply chain agents

4 agents · n8n · Owner Priya Raman

Critical
Runs
17,940
Success
95.6%
Duration
4.1 s
Spend
£3,180
Issues
2

GPT-6 Astra, Claude Opus 5, DeepSeek V4.1 (self-hosted)

Last run 2 min ago

Customer support agents

3 agents · n8n · Owner Tom Whitcombe

Healthy
Runs
24,600
Success
99.4%
Duration
1.2 s
Spend
£2,510
Issues
0

GPT-6 Astra, Gemini 3.8 Flash

Last run 40 s ago

Open the consoleAll six pages on sample data. Every figure is fictional.

How it works

Mirador runs as containers in your own cloud, beside your agents. Mirador reads what the tools your agents run on already record, and the analysis calls your own model account. Nothing crosses to us.

How Mirador runs: your agents, the tools they run on and Mirador inside your cloud; the analysis goes to your own model account; nothing goes to Mirador Systems.Your cloudYour model accountMirador SystemsReceives nothingAPIs and tracesYour agentsn8nLangGraphCrewAIOpenTelemetryMiradorContainersModel APIFor the analysisHow Mirador runs: your agents, the tools they run on and Mirador inside your cloud; the analysis goes to your own model account; nothing goes to Mirador Systems.Your cloudYour model accountMirador SystemsReceives nothingYour agentsn8nLangGraphCrewAIOpenTelemetryMiradorContainersModel APIFor the analysis

No SDK to write.

Mirador reads your agents through the APIs your tools already expose and receives OpenTelemetry traces from everything else. Mirador reads whichever model your agents call.

Run 8816 of PO matching, read from the trace. Read the PO took 4.0 s, Match to invoice took 1.1 s, Call OpenAI took 20.0 s, Write result took 0.3 s. Call OpenAI ran outside your cloud.

Security

The questions a security review asks first, answered in a line each.

Every answer in full, with the disclosure policy

Where does it run?
In your own cloud, as containers.
What leaves your network?
Nothing reaches Mirador. The analysis calls your own model account.
What is stored?
Redacted payloads only, in your own database.
How do I report a vulnerability?
Write to security@mirador.systems. The policy is on the security page.

Founder and design partners

Ahmed Abdelhedi

Five design partner places for Q4 2026. A free 90-day pilot on your own infrastructure, success criteria agreed at the start, and a direct line to the founder.

Talk to the founder

One flat monthly price per AI system, with no charge for traces, spans or seats.

An AI system is one business process run by one or more agents, such as your supply chain agents. Every agent and run inside it is included.

Starter

£299per month
Up to 5 AI systems
  • Health, explanations, cost and compliance
  • Two compliance packs
  • Three users
  • 30 day retention
  • Email alerts
Talk to the founder

Growth

£799per month
Up to 25 AI systems
  • Everything in Starter
  • All compliance packs
  • Ten users
  • 90 day retention
  • Microsoft Teams alerts
  • Engineer view
Talk to the founder

Prices exclude VAT. Every plan runs in your own cloud. Annual billing takes 20% off.

More than 25 systems, an air-gapped install, SSO or procurement paperwork? Talk to the founder.

Talk to the founder.

Tell us what you run and what you need to see. Ahmed reads every request and replies personally within one working day.

Or write to hello@mirador.systems.

For example, n8n and LangGraph on Azure.

Up to 500 characters.

Privacy. No mailing list: we send updates only if you tick the box.