Skip to content
AXLOOP AI
Menu
ProductArchitectureCompanyEarly access
Abstract network of distributed agent devices and telemetry paths

MCP Fleet Observability

Your agents are everywhere.Your telemetry is nowhere.

AxLoop AI instruments every MCP tool call on every device in your organization — laptops, phones, IDEs, embedded apps — and federates it into a single, locally deployable control plane.

See the product
EVERY CALLClient-edge spans, not server logs
LOCAL-FIRSTTelemetry never leaves the perimeter
SHADOW MCPSee servers nobody approved

01
The insight

The client edge is completely dark.

A developer runs a desktop AI client with a dozen MCP servers on a laptop. A colleague runs it on a phone. Three engineers have custom configs in their IDEs. Nobody can see what tools are being called, at what rate, whether any are failing, or what data is passing through them.

The blind spot

  • 01Which MCP tools are actually invoked — by whom, from which device
  • 02Latency breakdowns: the MCP server, the tool, or the upstream API?
  • 03Error rates and failure patterns across the fleet
  • 04Token and cost attribution per user, team, device and tool
  • 05Shadow MCPs — unauthorized servers on unmanaged devices
  • 06Data flow: sensitive information passing through tools it shouldn't touch
Unlit edge devices fading into darkness beside a brightly instrumented server hall
Illustration · Left: the uninstrumented edge · Right: where everyone is looking

02
The product

One dashboard. Every device. Real time.

Cross-device flamegraphs from the agent's reasoning step through the MCP call to the downstream API. Fleet rollups with per-device and per-user drilldown. Cost attributed to the team that spent it.

FLEET VIEW · ILLUSTRATIVE SAMPLE DATA128,400 MCP CALLS / 24H
DEVICES312+18 this week
ERROR RATE1.8%3 tools degraded
SPEND ATTRIBUTED$9,140across 14 teams

TOOL CALL STREAM

  • browser.navigatechrome · a.chen512 ms
  • internal.crm_lookupsidecar-07187 ms
  • shell.execmbp-14 · shadow-mcp73 ms
  • vector.searchandroid-14 · s.iyer264 ms
  • fs.read_filembp-14 · a.chen42 ms
  • postgres.queryios-17 · r.patel318 ms

LATENCY DISTRIBUTION

p50 88 msp95 640 msp99 1310 ms
CROSS-DEVICE TRACEagent.plan→mcp.tool_call→http.upstream→db.query

03
Architecture

Local-first. Federated. OTel-native.

No agent telemetry needs to leave the organization's perimeter. A regulated company cannot send agent telemetry to a SaaS — full stop. Local deployment is not a nice-to-have, it's the only viable enterprise sale.

PHONE APP
CLAUDE DESKTOP
DEV LAPTOP
BROWSER AGENT
↓   AXLOOP AGENT · SDK / DAEMON / OS HOOK   ↓
LOCAL OTEL COLLECTOR · PER DEVICE
ENCRYPTED · BATCHED SYNC ↓
FEDERATED AGGREGATION NODE · ON-PREM / VPCAXLOOP PLATFORM · traces · logs · metrics · per-device · fleet view
OpenTelemetry nativeW3C TraceContextRust · Swift · Python · Node · GoClickHouse · Tempo · LokiKubernetes or Docker Compose
Read the full architecture

04
Roadmap

Collector. Dashboard. Governance.

An open-source collector at the edge, a self-hosted fleet view on top of it, and a policy layer that turns visibility into control.

01
In development · Open source

The Collector

A zero-code instrumentation layer that wraps an MCP client and captures every tool call: name, redacted parameters, latency, success or failure, token usage. Stores locally, syncs encrypted batches.

  • Desktop client plugin
  • iOS + Android SDK
  • VS Code & JetBrains extension
  • Sidecar for embedded agents
02
Next · Self-hosted

The Fleet Dashboard

A self-hosted web UI on Kubernetes or Docker that turns raw spans into a fleet view — with drilldown all the way to a single device.

  • Cross-device flamegraphs
  • Fleet rollups + per-user drilldown
  • p50 / p95 / p99 by client type
  • Anomaly detection · cost attribution
03
On the roadmap · Planned

Policy & Governance

Observability becomes control: rules that decide which tools may run, on which devices, for which teams — with an audit trail behind every decision.

  • Restrict tools on unmanaged devices
  • Rate alerts per tool, per device
  • Shadow MCP detection
  • Audit log with conversation replay

05
Positioning

Everyone else is watching the servers.

ApproachWhere the edge stays dark
Server-side APMBuilt for backends. The MCP calls that happen on a laptop or phone never reach it.
Gateway control planesSee traffic that passes through the gateway. Direct client-side MCP usage stays invisible.
Vendor-native telemetryScoped to one vendor's own surfaces, not the mixed fleet most teams actually run.
Crash & error reportingReports app failures. No MCP semantics: no tool names, latencies, tokens or data flow.
AxLoop AIEvery MCP call on every device, normalized into one OTel schema. Local-first.

06 · Why now

MCP adoption moved faster than the tooling around it.

Agents went from demos to daily work in a single cycle, and they run on the devices people already carry. The instrumentation layer for that edge doesn't exist yet — so we're building it in the open.

About AxLoop