The instrument that explains what your agents did.
Mirador watches every AI agent and workflow you run, whatever you built it with, scores health and compliance, tracks the cost, and says why things happened. It runs in your own cloud.
A black hole seen from the edge of its accretion disc. Hot gas orbits the shadow, brighter on the side that moves towards you, and a thin ring of light hugs the horizon.
Request a demoAgents absorb data without limits. That is the problem.
A multi agent system will read anything it is given, send it anywhere it can, and cost whatever it costs, and none of that shows up until someone asks. Mirador is the instrument at the edge. It sees what goes in, what comes out, and what it cost, before anything crosses a line you did not draw.
A black hole seen from the edge of its accretion disc. Hot gas orbits the shadow, brighter on the side that moves towards you, and a thin ring of light hugs the horizon.
Event Horizon
the console. The boundary your agents do not cross without you seeing it.
Applications and the tools they call
Estate health
Spend by model, 30 days
- Claude
- GPT
- Gemini
Spend by flow
- £128
- £96
- £71
- Nine others£117
Runs by hour
Run 8811, candidate screening, 1.4 s
Compliance
- intervene: sends customer names to an external model. No approval recorded. GDPR art. 6
- attention: keeps prompts with personal data past the retention you set. ISO 42001 A.7
Events
- 14:32:07latency p95 1,840 ms, provider slower since Tue
- 14:31:52run 8812 ok, £0.04
- 14:31:40finding raised: external model, no approval
- 14:31:12run 8811 ok, 214 candidates ranked
- 14:30:58run 8810 ok, £0.02
Preview of the v1 console. Sample data.
Four things a leader needs to know about an application, on one page.
Health
Every flow gets a status. It comes from success rate, latency, cost per run, tool errors and the quality of what the agent produced, compared against its own history. You read one colour, then one sentence.
Latency p95 is 1,840 ms, up against its own last thirty days. Success rate and cost per run are unchanged.
Explainability
When something changes, Mirador says what changed and why. “Latency is up because the model provider is slower, not because your prompt changed” is a sentence a leader can act on. The engineer opens the run underneath it.
Cost per run fell 12 percent after the prompt change on Monday. Output quality is unchanged across 340 runs, so the saving is real.
Open the runCost
Spend by flow, by model and by month, with the trend. You see which flow carries the bill and whether that is rising. Prices come from the provider, so the number is the number.
- £128
- £96
- £71
- Nine others£117
Compliance
Mirador reads your workflows before it reads your traffic. A flow that sends personal data to an external model without an approval record lowers the score on day one. You choose the packs: GDPR, ISO 42001, or your own rules.
- intervene: sends customer names to an external model. No approval recorded. GDPR art. 6
- attention: keeps prompts with personal data past the retention you set. ISO 42001 A.7
No SDK to write.
Mirador reads workflows through the APIs your tools already expose and receives OpenTelemetry traces from everything else. n8n, LangGraph, CrewAI, the OpenAI Agents SDK, Google ADK, Pydantic AI, and anything that emits a trace. Statistical checks run on every execution. The plain English analysis runs through your own model account, so the data stays where it already is. Install as containers in your own cloud.
Six applications on three orbits around the Mirador mark, built with n8n, LangGraph, CrewAI, the OpenAI Agents SDK and Google ADK, each coloured by its status. Data flows inwards to Mirador, and out to Event Horizon, Teams alerts and the compliance report.
The person who signs it off, and the person who fixes it.
For the leader
A page you can read in thirty seconds. Health, spend, compliance, and the sentences that explain them. Alerts in Microsoft Teams or email when something needs a decision, and quiet when it does not.
For the engineer
Every run, every step, every model call. Prompt versions side by side. The anomaly the analysis found and how it reached that conclusion. The same data as the leader's page, with nothing hidden.
Flat monthly pricing. You pay for flows, not for traces, spans or seats.
Starter
- Health, explanations, cost and compliance
- Two compliance packs
- Three users
- 30 day retention
- Email alerts
Growth
- Everything in Starter
- All compliance packs
- Ten users
- 90 day retention
- Microsoft Teams alerts
- Engineer view
Enterprise: self hosted in your own cloud, unlimited flows, SLA and onboarding. Priced on request. Talk to us
Annual billing takes 20% off. Prices exclude VAT.
Who is building it.
Mirador Systems is a UK company. The team sits on the BSI committee that contributes to the ISO 42001 family of AI standards, has deployed multi agent systems inside automotive engineering, and published the research on multi agent observability that this product comes from, where the existing tools showed what happened and none of them said what it meant. We are onboarding a small number of design partners now.
Five things people ask first.
Does our data leave our environment?
No. Mirador runs as containers in your cloud, stores payloads in your own object storage, and runs the analysis through your own model account.
Which tools does it support?
Anything that exposes a workflow API or emits OpenTelemetry traces. n8n and LangGraph have the deepest integrations at launch. CrewAI, the OpenAI Agents SDK, Google ADK, Pydantic AI and custom agents connect through OpenTelemetry. Ask us about yours.
What does the analysis cost us in model calls?
Typically between £15 and £40 a month for ten flows on a small model, because statistical checks run on every execution and the model only reads a sample plus every failure. You control the model and the sample rate.
Is this an observability tool?
It sits on top of that layer. Observability tools show you traces. Mirador tells you whether the agents are healthy, what they cost, what went wrong and what to do, and keeps the traces underneath for the engineer.
When can we use it?
We are onboarding a small number of design partners now. Request a demo and we will show you the product on your own workflows.
See it on your own workflows.
Tell us what you run. We reply within one working day.