AI-native observability · MCP ready

Your AI.
Now on call.

Give your agent the production context it’s been missing. Trace a 500. Find the slow dependency. Investigate a memory spike. All from the conversation you’re already in.

You don’t even have to open a dashboard.

Connect your agent See it investigate
Setup target
About 5 minutes*
Your stack
Standard OpenTelemetry
Your control
Read-only MCP
Bring your own agent

Claude Cursor Codex+ any MCP client

Claude Code/ production
Scripted demo

Why is checkout returning 500s? Find the failing span.

Following the request through Lantern…

Lantern MCP → query_spans → get_trace
Database connection timeout in checkout-api. Inspect the connection pool and recent changes.

lantern · MCP · read-only

Real MCP tools. Illustrative data. Your agent stays in the driver’s seat.

01 / CONNECTYour telemetry. Your agent.Keep your OpenTelemetry stack. Bring the AI you already use.
02 / ASKQuestions in. Evidence out.Errors, traces, logs and metrics, connected through MCP.
03 / UNDERSTANDGet back to building.Find the failing call and decide what to do next.

The five-minute start

From “something’s wrong”
to your first real answer.

Let your agent handle the setup. Keep your instrumentation, connect your context, and ask the question that brought you here.

  1. 01
    ~1 MIN

    Connect your agent.

    One command adds Lantern’s hosted MCP server. Sign in once in your browser; your agent can only read.

  2. 02
    ~2 MIN

    Let it wire up your telemetry.

    Your agent gets the OpenTelemetry setup for your stack from Lantern, adds the exporter with your ingest key and checks that data arrives.

  3. 03
    ~2 MIN

    Ask your first production question.

    “Why is checkout failing?” Your agent follows the evidence. No dashboard tour required.

Hand this to your agentCOPY → PASTE → CONNECT
Help me connect this project to Lantern.

1. Add Lantern's read-only MCP server:
   claude mcp add --transport http --scope user lantern https://mcp.lanternstack.sh/mcp
   Then tell me to run /mcp and sign in to Lantern.

2. Call instrumentation_guide for this project and add
   the OpenTelemetry exporter it describes, using the
   LANTERN_OTLP_ENDPOINT and LANTERN_INGEST_TOKEN I export
   locally. Never ask me to paste keys.

3. Check setup_status until my services appear, then
   audit_telemetry. Then help me investigate my first
   production question.
Ask before deploying or restarting services.
Installation guide

*About five minutes with a Lantern account, an ingest key and existing OpenTelemetry instrumentation. Starting fresh? The installation guide walks through the prerequisites.

Prefer to wire it up yourself? Here’s the OpenTelemetry configuration. 
exporters:
  otlphttp/lantern:
    endpoint: https://ingest.lanternstack.sh   # /v1/traces is appended
    compression: gzip
    headers:
      Authorization: "Bearer ${env:LANTERN_INGEST_TOKEN}"
processors:
  batch:
    send_batch_max_size: 256       # tune so encoded requests stay below 4 MiB
service:
  pipelines:
    traces:
      receivers: [otlp]
      processors: [batch]
      exporters: [otlphttp/lantern]
  1. Get an ingestion keySeparate from read access: ingestion keys can’t query data.
  2. Add the exporterCollector, SDK environment variables or the operator’s Instrumentation resource.
  3. Ask your agentVerify the connection with get_status, discover services and read your first trace.

Production context, on tap

Your agent already knows your code.
Now it can see what happened.

A stack trace is a clue. The request, its dependencies, the logs around it and the metrics at that moment tell the rest of the story. Lantern puts that context one MCP call away.

Explore the MCP integration

Read access. You keep control. Lantern’s MCP tools investigate your telemetry. Your agent proposes the next step; you decide what changes.

Follow the request

Find the failing span and the call it depends on.

query_spans → get_trace
TRACES

Read what happened around it

Connect an exception or timeout to its trace.

query_logs → stream_logs
LOGS

Check the bigger picture

Compare latency with the memory metrics you export.

list_metrics → query_metrics
METRICS
One MCP connection. Your agent of choice.

The Lantern dashboard

Your whole system.
One clear picture.

Zoom out to the overview. Zoom in to the exact span. Explore your production data in a dashboard built for curious humans, with the same context your agent sees.

app.lanternstack.shINTERACTIVE PREVIEW

Workspace telemetry Overview

A pulse on production.

production

Requests over time

SpansErrors

Duration heatmap

duration / time

Wide events 6 sample spans

Select a row
Synthetic request spans. Select an operation to inspect its fields.
Time UTCServiceOperationDurationStatus
Synthetic telemetry · no workspace connected14:15–14:30 UTC

An interactive look at Lantern. Illustrative data, generated in your browser.

01 / SEE THE PATTERN

Start with the big picture.

Request volume, errors and latency, side by side. Your overview gives the investigation a starting point.

02 / FOLLOW THE EVIDENCE

Every event has a story.

Inspect its fields, follow the trace and find the dependency that took the time.

03 / WORK YOUR WAY

A board. A query. A conversation.

Keep exploring in the dashboard or take the question to your agent. The context comes with you.

Open the dashboard Explore the query playground

Query playground

Where, summarize, by. That’s the whole language.

These are the same filters your agent uses through MCP. Explore 6,000 synthetic checkout events here, then take the question back to your conversation.

where
summarize
by 

            

            
            

            

          

        

      

    




    

      

        

          
Under the hood

          
Small, honest, exact.

        

        

          
Questions after ingestion
Keep the context. Decide what to group by when you are investigating, not when you are instrumenting.

          
Aggregates in the engine
Counts, errors and percentiles are computed over the whole window, not a sampled page of rows.

          
Original OTLP kept
Every record keeps its original protobuf next to typed, queryable attributes and nanosecond timestamps.

          
Exact metrics
Raw, latest, increase, rate, delta, distribution and summary, with unsupported operations failing loudly.

          
Separate keys
Ingestion keys cannot read. Read keys cannot ingest. The workspace is chosen by the server, not the client.

          
Encrypted at rest
TLS end to end through Cloudflare, encrypted storage and encrypted backups.

        

      

    


    

      

        

          
FAQ

          
Good questions.

        

        

          
What is a wide event?
One record per unit of work that carries all of its context: route, user plan, region, build, feature flags, database statement. In OpenTelemetry terms, a span with rich attributes. Because nothing is pre-aggregated, you can ask new questions later.

          
Do I need a new SDK or agent?
No. Lantern accepts standard OTLP over HTTP with protobuf and gzip. Keep your OpenTelemetry SDKs and Collector, and add an exporter with your ingestion key.

          
Do I have to use the dashboard?
No. Once telemetry and MCP are connected, your agent can discover services, query traces and logs, inspect spans and read metrics directly. The dashboard is there whenever you want a visual view.

          
Which agents can I use?
Any client that supports remote MCP servers, such as Claude Code, Codex or Cursor. Add https://mcp.lanternstack.sh/mcp and sign in once with your Lantern account. Lantern supplies the production context; your chosen agent handles the conversation. See the MCP setup guide.

          
Which signals are supported?
OpenTelemetry traces, metrics and logs over OTLP/HTTP with protobuf and gzip. Host and pod memory metrics need to be exported by your instrumentation or Collector; they are not measured automatically from a span.

          
How does the AI integration work?
Lantern hosts a read-only MCP server with bounded tools such as setup_status, query_spans, search_logs and get_trace. You approve your agent once in your browser, and you can disconnect it at any time. Your agent runs the same queries you see in the app. Telemetry is always treated as data, never as instructions.

          
Is this another SigNoz or Datadog?
No. Lantern does not start from fixed APM screens or per-metric definitions. It starts from events and lets you slice them any way, in the spirit of Honeycomb and Axiom, with an interface that aims to be friendlier and AI-first.

          
How much does Lantern cost?
The launch plan starts at $30 USD per month, including 1 TB of ingest, 100 GB-hours of query compute and 100 GB of storage. Usage above those allowances costs extra. Explore the pricing page and workload planner.

          
Can I use it today?
Lantern is in early access with a small number of teams. Sign in if you have an account, or ask us for one.

        

      

    


    

      

        
        
Your next incident
could start with a question.

        
Bring your telemetry. Bring your agent.
Lantern brings the production context. Now in early access.

        
Read our terms of service, privacy notice and refund policy.

        

          Request early access
          Read the setup guide