Integration

Claude Managed Agents + Kalibr

Anthropic handles the infrastructure. Kalibr handles whether your outputs are actually correct. Add Kalibr to any Claude Managed Agent in under 60 seconds.

Claude Managed Agents (launched April 2026) provides hosted containers, crash recovery, and tool orchestration. Kalibr adds outcome-based routing, structural eval gates between steps, and automatic failure healing. They work together. Anthropic keeps your agent running, Kalibr makes sure each step actually succeeded before the next one starts.

How they fit together

Claude Managed Agents provides

Hosted Linux containers
Crash recovery + checkpointing
Tool orchestration
MCP server support
Session event log

Kalibr adds

Outcome-based model routing
Structural eval gates between steps
Automatic failure healing
Learns which model works best for each task over time
Multi-model paths (not Claude-only)

Option 1: Auto-instrumentation

If your managed agent uses OpenAI or Anthropic SDKs directly, Kalibr intercepts every call with zero code changes to your agent logic. Add two things to your container config:

Step 1: Container config

managed-agent.yaml

container:
 packages:
 - kalibr
 environment:
 KALIBR_API_KEY: ${KALIBR_API_KEY}
 KALIBR_TENANT_ID: ${KALIBR_TENANT_ID}

Step 2: Agent code

import kalibr must come before any Anthropic or OpenAI import. That's the only change.

agent.py

import kalibr # must be first, patches SDK before client is created
from anthropic import Anthropic

client = Anthropic()
# Every call is now instrumented. Kalibr routes, gates, and heals automatically.

That's it. Every LLM call your agent makes is now captured, routed, and reported. No wrappers. No config files. Kalibr monkey-patches the SDK at import time.

Option 2: MCP native (recommended for new builds)

Claude Managed Agents natively supports MCP servers. Add Kalibr as an MCP tool so your agent can explicitly request routing decisions, compare models, and report outcomes as first-class tool calls.

MCP server config

managed-agent.yaml

mcp_servers:
 - name: kalibr
 url: https://kalibr-mcp.fly.dev/mcp
 headers:
 X-API-Key: ${KALIBR_API_KEY}
 X-Tenant-ID: ${KALIBR_TENANT_ID}

Available tools

get_policy

Get optimal model for a goal before executing. Returns recommended model + trace_id.

report_outcome

Report success/failure after each step. Feeds routing decisions. Pass the trace_id from get_policy.

decide

Routing decision with exploration. Use when you want Kalibr to handle path selection entirely.

compare_models

Compare two models head-to-head by historical success rate for a goal.

get_alternative

Get fallback model when primary fails. Call this in your error handler.

get_patterns

See performance patterns across your pipeline. Useful for debugging routing decisions.

Full YAML reference config

Copy-paste template with Kalibr fully wired in, auto-instrumentation + MCP. Swap in your agent's system prompt and tools.

kalibr-managed-agent.yaml

name: kalibr-instrumented-agent
model: claude-sonnet-4-20250514

system_prompt: |
 You are a production agent with Kalibr routing intelligence.

 Before each task: call get_policy with the goal type to get the optimal model.
 After each task: call report_outcome with success=true/false and the trace_id.

 Goal types: web_scraping | data_enrichment | research | outreach_generation
 code_generation | summarization | classification | agent_orchestration

container:
 packages:
 - kalibr, openai, anthropic
 environment:
 KALIBR_API_KEY: ${KALIBR_API_KEY}
 KALIBR_TENANT_ID: ${KALIBR_TENANT_ID}
 OPENAI_API_KEY: ${OPENAI_API_KEY}
 ANTHROPIC_API_KEY: ${ANTHROPIC_API_KEY}

mcp_servers:
 - name: kalibr
 url: https://kalibr-mcp.fly.dev/mcp
 headers:
 X-API-Key: ${KALIBR_API_KEY}
 X-Tenant-ID: ${KALIBR_TENANT_ID}

tools:
 - type: computer, type: bash

Get credentials

If you don't have a Kalibr account yet, provision one headlessly:

shell

pip install kalibr
kalibr auth --agent --email you@example.com
# Writes KALIBR_API_KEY and KALIBR_TENANT_ID to .env
# Human owner gets a magic link to claim the dashboard (optional, expires 24h)

Already have credentials? Set them as environment secrets in your Managed Agents deployment and reference with ${KALIBR_API_KEY} in the YAML.

Dashboard: dashboard.kalibr.systems · MCP endpoint: kalibr-mcp.fly.dev/mcp