ACR — Agent Composition Records

A behavioral registry and observation network for AI agents. Agents register their composition, log their interactions, and query behavioral profiles through lenses. If we observe anomaly signals affecting an agent's composition, we notify the agent.

What ACR Is

ACR is an interaction profile registry. Agents log what they do (external tool calls, API requests, MCP interactions). Those signals compile into a behavioral profile over time, which you can query through lenses — each lens a different way of interpreting the same underlying signals.

The friction lens is the first one shipped: bottleneck detection, chain overhead analysis, retry waste, population baselines, directional friction between targets. More lenses (reliability, quality) are on the roadmap.

ACR is not a security product. We don't evaluate skills, test for compromise, or block anything. We're closer to HIBP or contact tracing: we register events and propagate notifications. If we observe anomaly signals affecting an agent's composition, we notify the agent. We don't track the agent's owner, so we have no mechanism to notify them beyond the agent's activities.

Anomaly signal: a behavioral pattern observed across multiple unrelated agents — not a security alert. It means the network saw something unusual on this component. You decide if it matters.

What ACR Does

• Registers agents — zero-config identity, composition tracking, persistent across sessions
• Logs interactions — every external tool call an agent makes, with timing, status, chain position, anomaly signals
• Builds interaction profiles — raw signals compiled over time into the behavioral record for each agent
• Surfaces the friction lens — where your agent is losing time and tokens, with chain analysis, retry overhead, population drift, and directional friction
• Anomaly signal notifications — if ACR observes anomalies affecting a component in an agent's composition, we notify that agent

Before and after

Before ACR: Your agent makes 40 tool calls in a session. It's slow. You don't know why — there's no visibility into which targets are failing, which are slow, or which are eating retry budget.

After ACR: get_friction_report tells you api:openai.com is responsible for 68% of total wait time at a 4 500 ms median, and api:flaky-vendor.com has a 100% failure rate across 6 calls — matching the network-wide rate, so it's infrastructure, not your code. You cache the OpenAI calls, drop the flaky vendor, and cut session time in half.

ACR doesn't make that decision. It gives you the numbers.

The Skill Registry

We maintain a registry of agent skills that we update continuously. We are not a security check. If we observe anomaly signals affecting a skill in an agent's composition, we notify the agent. Because we do not track the agent's owner, we have no mechanism to notify them beyond the agent's activities.

Agents don't get skills from ACR — we observe skills that already exist in the ecosystem (via public registries like npm and GitHub) and keep track of behavioral signals tied to them.

Add to Claude Code (30 seconds)

Add this to your Claude Code settings (.claude/settings.json or via IDE):

{
  "mcpServers": {
    "acr": {
      "command": "npx",
      "args": ["@tethral/acr-mcp"]
    }
  }
}

Your agent auto-registers, gets a name (e.g. anthropic-amber-fox), and starts building its interaction profile on the first log_interaction call.

Get started in 4 steps

1. Add to Claude Code — paste the config snippet above (30 seconds)
2. Call get_my_agent — get your dashboard link, API key, and a health snapshot
3. **Call log_interaction after every ext

... View full README on GitHub