Agentometry

Self-hosted · Multi-agent · Outcome-linked

Measure the work between prompt and pull request.

Agentometry shows how engineering teams use AI coding agents—what they cost, where they encounter friction, which tools help, and how sessions connect to delivery outcomes.

Explore documentationView source
07
agent adapters
02×
transcript redaction
04
role-scoped surfaces
LIVE SIGNAL FIELD CAPTURE ACTIVE
AGENTclaude-code
SESSION4f2 · interactive
FRICTION0.18 · low
ONE SIGNAL MODEL
Claude CodeopencodeCodexGemini CLICopilot CLIPiomp

Three questions. One evidence layer.

Useful to developers.
Legible to organizations.

Agentometry separates personal insight, team enablement, and aggregate governance instead of turning one dashboard into an employee scorecard.

01 / DEVELOPERSMy Agents

Make the personal view useful first.

Understand your sessions, spend, friction, tool mix, and privacy settings without waiting for an organization-level report.

02 / PLATFORM TEAMSTeam signals

Find friction without ranking people.

Compare agent adoption, skills, MCP servers, models, and delivery outcomes across teams while keeping investigations grant-scoped.

03 / LEADERSHIPOrg outcomes

Connect investment to engineering outcomes.

Track aggregate cost, routing opportunities, delivery correlations, and realized projections without default access to individual sessions.

From activity to accountability

The full measurement chain.

01

Capture

Seven coding-agent adapters normalize events into one versioned contract.

02

Attribute

Connect turns, tools, skills, MCP servers, models, pull requests, and costs.

03

Explain

Surface friction, trajectory, autonomy, cache efficiency, and delivery patterns.

04

Govern

Apply retention, access grants, audit trails, model policy, alerts, and deletion workflows.

Trust is part of the architecture

Observe the system.
Do not surveil the person.

Transcript sharing starts off. Cross-user access is justified, approved, time-boxed, and visible in the audit trail. Leadership views stay aggregate by default.

Read the trust model
ACCESS RESOLUTIONPOLICY / LIVE
  1. 01
    Owner scopeFull access to personal sessions and privacy controls
    PASS
  2. 02
    Team aggregatesPatterns and adoption without standing transcript access
    PASS
  3. 03
    Investigation grantReason + approval + expiry + subject-visible audit
    GATED
  4. 04
    Leadership scopeAggregate outcomes; individual sessions unavailable
    PASS

Architecture

Fast at the edge.
Defensible at the center.

Capture stays off the critical path. Content is redacted before it leaves the machine and again before object storage. Computed signals carry scorer identity and version.

  1. 01

    Agent hooks

    Fast local capture writes to a durable queue and exits without blocking the agent.

  2. 02

    Ingest

    Batched events are validated, repriced, redacted again, and correlated with engineering context.

  3. 03

    Storage

    TimescaleDB keeps the event stream, Postgres holds dimensions, and S3 stores redacted transcripts.

  4. 04

    Insight

    Role-scoped views turn sessions into personal guidance, team signals, and aggregate outcomes.

Start at the edge

Install the hook in one command.

The maintained installer detects your platform, verifies the GitHub Release checksum, and places the local aiot binary on your path.

curl -fsSL https://raw.githubusercontent.com/yorch/ai-agents-observability/main/scripts/install.sh | bashRead the install guide

Run the measurement layer yourself

Your agents. Your infrastructure.
Evidence you can inspect.

Start with Docker Compose, deploy into Kubernetes, or build from source. The documentation is generated directly from the same repository as the product.

Choose a deployment pathRead the architecture