Komodor vs Metoro (2026)

Agents act on Kubernetes and cloud infrastructure through the platform's tools, MCP gateways and integrations, with ready to run workflows for incident troubleshooting, CI/CD remediation and cost optimization across clusters.

Catalog agents cover AWS, Azure and Google Cloud with read only access, Datadog and Grafana across traces, logs, metrics and profiles, Kubernetes resources and clusters including GPU monitoring, and pull request review. Credentials are supplied once, encrypted in a secure store and handed to each run rather than built into images. Teams expose their own tools to agents through the MCP Gateway.

SourceKomodor, komodor.com and docs.komodor.com agent catalogread 2026-10-05

Guardian inspects code changes and creates pull requests through the GitHub integration and posts findings to Slack, beside PagerDuty, Rootly, webhook and event ingestion integrations and its own eBPF, OpenTelemetry and Prometheus telemetry.

Sourcemetoro.io/docs/ai-sre/overviewread 2026-09-29

The Agentic Operations Platform turns skills, scripts or runbooks into production agents and runs agent fleets built on LangChain, Google ADK, the Claude Agent SDK or AgentCore beside Komodor's own, under shared governance, with ready to run workflows. A catalog Orchestrator agent routes work to other agents in the fleet, and agents can also call one another, run as steps in a workflow, or start from chat, channels, modules or a manual run.

The Incident Management module's Production Alert Investigation workflow collapses repeated and related alerts into one incident, sends observability, Kubernetes and cloud investigators to gather evidence in parallel, writes a root cause conclusion with the evidence attached, and proposes a fix that waits for approval before anything is applied.

Once a decision and an approval are behind it, a Remediation Executor agent carries the fix out, acting on the live resource or opening the pull request that fixes the source, and it cannot answer an approval itself. Existing agents can be imported without a rewrite, and the platform page adds Codex to the supported frameworks. Underneath sit a control plane that Komodor operates, an SDK for building agents, and workers that run in Komodor's infrastructure or inside the customer's network.

SourceKomodor, komodor.com/platform and docs.komodor.com overview, agent catalog, Remediation Executor and Incident managementread 2026-10-25

Customers write runbooks for each alert in natural language, as numbered investigation steps Guardian follows when the alert fires, so the customer edits the flow. Guardian also chains deployment verification, root cause and fix pull requests. Metoro does not describe branching or conditions.

Sourcemetoro.io/docs/ai-sre/runbooksread 2026-09-29

Alert Intelligence and Incident Management modules work alerts and incidents as they arrive, and change intelligence and proactive reliability optimization run continuously across clusters. Runs start from alerts, schedules, chat, channels, pull requests, workflow steps, other agents and manual runs. Schedules use standard cron expressions with a timezone and optional input, and fire only while both the schedule and its agent are enabled. Inbound endpoints let outside systems start a run with an HTTP request.

SourceKomodor, komodor.com/platform/pricing-and-plans, /platform and docs.komodor.com Triggers and schedulesread 2026-10-05

Guardian starts on its own on a new_deployment event, on anomalies it detects such as error rate spikes, and when an alert fires from a known threshold or condition, and runbooks run when a configured alert triggers.

Sourcemetoro.io/docs/ai-sre/overviewread 2026-09-29

Agents run on 'a living knowledge graph' of the customer's Kubernetes and cloud estate, shared across the fleet. The graph relates services, owners and dependencies, and beside it sits a knowledge base of runbooks, postmortems, service docs, policies, decisions, guides and incident write ups that teams add as Markdown, text or PDF files, up to 20 files per upload and 10 MB each.

Each document is indexed for both semantic and keyword search and split at its headings, so an agent's search tool returns five matching passages by default and up to ten, each with the document, section and chunk it came from so a citation can be checked. Documents carry service, environment, type, owner and sensitivity tags, stay inside one account, and catalog agents search them by default.

SourceKomodor, komodor.com and docs.komodor.com Knowledge baseread 2026-10-25

Guardian investigates over Metoro's own persisted store of the customer's eBPF logs, traces, metrics, profiles and Kubernetes resources, with a service map and deployment history, compared before and after each rollout.

Sourcemetoro.io/docs/ai-sre/overviewread 2026-09-29

Agents work from shared memory: 'Every agent draws on the same shared memory and a living knowledge graph, so nothing starts from zero'. Memory belongs to each agent, and agents that share a space can read each other's learnings. After a run the agent reviews its own work and saves what is worth keeping as a technique, pointer, constraint, fact, feedback or preference, labeled by topic.

New memories wait as drafts for a reviewer agent, active memories stay retrievable until they are archived or folded into a better one by a coherence sweep, and archived memories stay readable for 90 days before they are permanently deleted. People filter, edit and archive memories in Manage memories, and an Activity view tracks how often they help.

SourceKomodor, komodor.com and docs.komodor.com Memoryread 2026-10-05

There is no memory the agent reads across runs. The Issues Catalog tracks recurring issues with the root cause, symptoms, impact and resolution steps, which is a record of the customer's incidents, not agent memory.

Sourcemetoro.io/docs/ai-sre/overviewread 2026-09-29

Higher risk actions are held for a person ('Maintain full control over higher-risk actions'), with agent RBAC, input, output and tool guardrails, budget caps and rate limits on every plan. Holds come from workflow steps written to pause for a person and from guardrails set to Hold for approval on a specific tool call, and each grant is single use and covers that exact call with its arguments. Pending decisions collect in the Attention inbox, with approve and deny beside each one, and the platform overview counts them on an Approvals card. Only a person can answer, in the console.

An agent cannot approve its own held action, and a rejection ends that path of the run. Every decision and the guardrail verdict behind it land in the run's evidence. Nothing pushes a pending approval to the approver yet, though a Slack thread now says when a run is waiting on a person. Behind a feature flag, an approver allowed to change policy can approve and stop asking, which edits the agent's action policy for that tool and is recorded in the audit trail.

SourceKomodor, komodor.com/platform/pricing-and-plans, docs.komodor.com Approvals and Attention, and the September 20 changelogread 2026-10-25

Guardian asks for approval before it takes an action, and a webhook event is sent when it does. Code fixes arrive as pull requests for review.

Sourcemetoro.io/docs/api-reference/webhooks/agent-action-approval-requestedread 2026-09-29

RBAC, agent RBAC and SSO come on every plan, with comprehensive audit logs and data retention controls. SOC 2 Type II and GDPR are listed on both plans.

Seven built in roles run from Owner and Admin through Developer, Operator, Member and Viewer to an Agent role for agent principals, and custom policies and roles can be scoped account wide, by name pattern, by label or with an expiry. Access is denied unless a grant matches, and an access explorer shows the effective permissions of any identity before a change goes live.

Agents carry their own identity, credentials never appear in run evidence, and knowledge base documents never cross account boundaries.

SourceKomodor, komodor.com/platform/pricing-and-plans and docs.komodor.com Roles and permissionsread 2026-10-05

SSO and SAML, RBAC and audit logs come with the Scale and Enterprise tiers, with setup pages for RBAC, SSO, SAML and API keys. Every tier is SOC 2 Type II certified.

Sourcemetoro.io/pricingread 2026-09-29

Komodor keeps a 'Full audit trail. Every verdict (allowed, redacted, blocked, or held) lands on the record, for every agent and every run'. Token usage is visible per agent.

Each run records a transcript, a trace timeline of reasoning, model calls and tool calls with their durations, logs, structured and readable output, artifacts, the skills loaded and the guardrail decisions. Token use is split into four classes with cost attached, and evidence stays readable for failed runs.

The History ledger filters runs by type, status and agent, the same records come back over the API, and the docs cover OpenTelemetry export.

SourceKomodor, komodor.com and docs.komodor.com Runs and evidenceread 2026-10-05

Each investigation produces an investigation document with all gathered data, and findings come with supporting evidence. That is evidence for each run, but there is no record of the queries, tools or steps Guardian took. Metoro's telemetry of the customer's cluster is what the product does, not a record of the agent.

Sourcemetoro.io/docs/ai-sre/alert-investigationsread 2026-09-29

Komodor offers to 'Run wherever your infrastructure lives: self-hosted on cloud or on-prem, or Komodor cloud', which puts the platform in the customer's own environment or in Komodor's cloud. Agents can run in Komodor's infrastructure when they only need public APIs, or as self hosted agents in the customer's clusters or VMs that keep one outbound HTTPS connection to the hosted control plane and need no inbound firewall rules or VPN.

The whole control plane can also be self hosted from a single Helm chart bundle, in beta and arranged with the account team, and then the only outbound traffic goes to the model and the systems the customer connects. The platform page lists self managed Kubernetes, EKS, GKE, AKS, serverless and on premises targets, and hosted and self hosted agents can sit in one fleet.

SourceKomodor, komodor.com, /platform and docs.komodor.com Deployment methodsread 2026-10-05

Three deployment models are offered, Metoro Cloud, Metoro BYOC and Metoro On-Prem, and there are on premises installation guides for the Metoro Hub (On-Premises 11.x).

Sourcemetoro.io/pricingread 2026-09-29

Komodor ships ready to run solution modules, each for its own job: Incident Management and Troubleshooting, Alert Intelligence, Proactive Reliability Optimization, Kubernetes and Cloud Cost Optimization, Change Intelligence and Risk Control, CI/CD Health and Remediation, and Production Readiness and Standards. Observability Cost Optimization is added on Enterprise.

The agent catalog is curated per account and its agents deploy as they are once credentials are supplied, covering cloud providers, APM tools, Kubernetes, code review, an orchestrator, remediation with a separate remediation executor, and an AWS cost analyzer. The platform page counts more than 50 agents across the modules.

SourceKomodor, komodor.com/platform/pricing-and-plans, /platform and docs.komodor.com agent catalogread 2026-10-05

Deployment verification, anomaly detection and root cause, alert investigations, runbook following and code fixes are capabilities of one agent, Guardian, not separately adoptable agents. The dashboard templates are not agents either.

Sourcemetoro.io/docs/ai-sre/overviewread 2026-09-29

The platform runs 'on your cloud, on-prem, or Komodor Cloud, with any model', with model routing that sends simple steps to cheaper models for token efficiency. Models come three ways.

Komodor can supply managed models with budgets it sets, a customer can register its own accounts with any of 18 providers, among them OpenAI, Anthropic, Azure, Bedrock, Vertex AI, Gemini, Mistral, Groq, DeepSeek, OpenRouter, Together AI, xAI, Cohere, Ollama, Baseten and Nebius Token Factory, plus hosted vLLM endpoints, and self hosted agents can send requests straight through the customer's own LLM gateway. Self hosted agents pick their models when they are built or deployed, and every model sits in one table whichever way it arrived.

SourceKomodor, komodor.com, /platform and docs.komodor.com Providersread 2026-10-05

The model is Metoro's choice, named as Claude Opus 4.5, a single named maker. Customers may optionally run Guardian's model calls on their own AWS Bedrock account, and otherwise Metoro's default credentials are used. That decides where the calls run, not which model maker serves them.

Sourcemetoro.io/docs/ai-sre/aws-bedrock-setupread 2026-09-29

Agents are built with MCP gateways and integrations that connect them to tools. Komodor's own REST API sits at /api/v1 on the account host and takes bearer API keys issued to service accounts or to people, and an interactive reference generated from the live service sits at /api/docs on the same host behind the account login.

The platform's MCP server at /internal/mcp checks every tool call against the same permissions as the HTTP API, with toolsets that search the platform, create and change agents, schedules and integrations, run agents, manage incidents, workflows and the gateway, and handle golden test scenarios. Claude Code, Codex CLI and Cursor are named as clients, and the pricing table lists API, MCP and infrastructure as code access on both plans.

SourceKomodor, docs.komodor.com APIs and MCP server, and komodor.com/platform/pricing-and-plansread 2026-10-05

There is an API with an OpenAPI spec (query logs, create environments, create or update alerts and dashboards) and webhook events. The Metoro MCP server runs on the workspace URL with API key auth and offers tools such as create_investigation and create_ai_issue.

Sourcemetoro.io/docs/integrations/metoro-mcp-serverread 2026-09-29

An Optimize pillar promises continuous improvement of agent quality, speed and cost. Evaluation rules send a share of an agent's runs, from 10 to 100 percent, to a panel of judge agents that score each run from 0 to 1 against named rubrics such as quality, precision and recall, with the reasoning attached to every verdict.

Model Evaluation grades shadow candidates on production traffic without touching the live agent's averages, and golden test scenarios and suites can be managed from the platform's MCP server. People can add their own star ratings, which stay separate from the judges' scores. The pricing table lists an out of the box eval framework, a golden scenario builder and per tool token evaluation on both plans, with custom evals on Enterprise.

SourceKomodor, docs.komodor.com Evaluations and MCP server, komodor.com/platform and /platform/pricing-and-plansread 2026-10-05

Deployment verification compares telemetry, log patterns and latency before and after each rollout, which validates the customer's change, not Guardian's own behavior. There is no evaluation of the agent's investigations or fixes.

Sourcemetoro.io/docs/ai-sre/overviewread 2026-09-29

Agents operate Kubernetes and cloud through tools, MCP gateways and integrations. No browser, desktop or computer control is documented. Fixes are carried out by a dedicated Remediation Executor agent working through the customer's own tools.

SourceKomodor, komodor.com and docs.komodor.com agent catalogread 2026-10-05

There is no browser, desktop or computer control. Guardian works through Metoro's telemetry, GitHub and Slack integrations.

Sourcemetoro.io/docs/ai-sre/overviewread 2026-09-29