🛡️ ocm-mcp-server

📖 Read the docs site → ocm-mcp-server.github.io

Star us ❤️ → 

AgentOps for Kubernetes fleets, done safely.

An MCP server that lets AI agents operate a multi-cluster Kubernetes fleet through an Open Cluster Management hub, with policy, approval, and audit between the model and your clusters.

The agent never holds a kubeconfig. Every write is policy-checked, human-approved, and traced.

OpenSSF Scorecard

📦 Get it  ·  ✨ Why  ·  🔌 Connect your agent  ·  🧭 Architecture  ·  🧰 Toolsets  ·  🛠️ Tools  ·  💬 Prompts  ·  🔭 Observability  ·  🚀 Quickstart  ·  📖 Wiki  ·  📚 Docs

_{The whole safe-remediation loop: investigate with free reads, propose a change, get rejected by the guardrails and correct it, wait for a human-signed token, apply, verify, and report from the audit log.}

Where to get it, and how it's vetted

• 📦 PyPI - ocm-mcp-server - pip install ocm-mcp-server (or run directly with uvx ocm-mcp-server). Every release is published straight from CI via OIDC trusted publishing - no long-lived tokens anywhere.
• 🗂️ Official MCP Registry - listed as io.github.ocm-mcp-server/ocm-mcp-server, so any MCP client or platform that browses the registry can discover and auto-configure this server (package, transport, and required env vars are all in the listing); the registry validates the listing against this repo and the PyPI package.
• 🐳 Container image on GHCR - docker run ghcr.io/ocm-mcp-server/ocm-mcp-server (kubeconfig mount shown in the deployment guide); built in CI with an SBOM and SLSA provenance attached, vulnerability-gated with Trivy, and signed keyless with Cosign so you can verify what you run.
• 🛡️ OpenSSF Scorecard - the repo's supply-chain security posture (pinned dependencies, branch protection, signed releases, ...) is scored automatically every week and published for anyone to inspect.

Why this exists

Your team runs many Kubernetes clusters. Sooner or later somebody asks the question: can an AI agent take the 2 a.m. page?

The quickest way to find out is to hand a model kubectl with cluster-admin and watch. In production that experiment ends badly, for three separate reasons:

• The model is non-deterministic. The same alert can produce a careful diagnosis one run and a kubectl delete the next.
• The credentials are real. There is no dry run between the model's decision and your production cluster.
• There is no record. When something breaks, you cannot reconstruct what the agent did, in what order, or on whose authority.

This project starts from a different observation: fleets already have a control point that humans trust every day, the multi-cluster hub. Open Cluster Management (a CNCF project) gives every fleet an inventory (ManagedCluster), a scheduler (Placement), and a delivery channel (ManifestWork). ocm-mcp-server exposes that hub to agents as a small set of typed MCP tools, and puts four independent layers between the model and your clusters:

None of these layers live in the system prompt, so none of them can be talked out of.

_{A write is two calls with a person between them. The token is bound to one content hash and one operation, it expires on its own, and offered a second time it is refused.}

_{Two writes, the same four gates. The privileged, unpinned one dies at Layer 1 and never reaches a cluster; the compliant one waits for a person to sign the exact content, then lands and is verified.}

_{A fleet operator's day with Claude, live from a cold start: install from PyPI, claude mcp add, inventory the fleet, reason about placement — then ship a new service the gated way: the privileged :latest shortcut is refused, the pinned proposal is signed by a human, applied with the token, verified, and the whole day is read back from the audit trail. — narrated MP4 · terminal cast.}

_{The same day, driven by four different agents. Identical ten chapters, identical server - only the agent asking changes, which is the whole point of speaking MCP rather than shipping a client. Codex · Gemini (Antigravity CLI). Re-record any of them with hack/demo-record.sh all.}

Connect your agent - any MCP client works

The server speaks standard MCP over stdio; nothing here is specific to one vendor's agent. That claim is demonstrated, not asserted: the same ten-chapter operator session is recorded against four different agents - Claude Code, Codex, Gemini through the Antigravity CLI, and IBM Bob Shell - driving the same server against the same fleet, each one really calling the tools, hitting the guardrail refusal, and applying only with a human-signed token. Re-record any of them with hack/demo-record.sh all.

Ready-made configs live in examples/ - see the index for where each file goes:

{
  "mcpServers": {
    "ocm-fleet": {
      "command": "ocm-mcp-server",
      "env": {
        "OCM_MCP_HUB_CONTEXT": "kind-hub",
        "OCM_MCP_SPOKE_CONTEXTS": "cluster1=kind-cluster1,cluster2=kind-cluster2,cluster3=kind-cluster3"
      }
    }
  }
}

{
  "servers": {
    "ocm-fleet": {
      "type": "stdio",
      "command": "ocm-mcp-server",
      "env": {
        "OCM_MCP_HUB_CONTEXT": "kind-hub",
        "OCM_MCP_SPOKE_CONTEXTS": "cluster1=kind-cluster1,cluster2=kind-cluster2,cluster3=kind-cluster3"
      }
    }
  }
}

[mcp_servers.ocm-fleet]
command = "ocm-mcp-server"

[mcp_servers.ocm-fleet.env]
OCM_MCP_HUB_CONTEXT = "kind-hub"
OCM_MCP_SPOKE_CONTEXTS = "cluster1=kind-cluster1,cluster2=kind-cluster2,cluster3=kind-cluster3"

{
  "mcpServers": {
    "ocm-fleet": {
      "command": "ocm-mcp-server",
      "env": {
        "OCM_MCP_HUB_CONTEXT": "kind-hub",
        "OCM_MCP_SPOKE_CONTEXTS": "cluster1=kind-cluster1,cluster2=kind-cluster2,cluster3=kind-cluster3"
      }
    }
  }
}

{
  "mcpServers": {
    "ocm-fleet": {
      "command": "ocm-mcp-server",
      "env": {
        "OCM_MCP_HUB_CONTEXT": "kind-hub",
        "OCM_MCP_SPOKE_CONTEXTS": "cluster1=kind-cluster1,cluster2=kind-cluster2,cluster3=kind-cluster3"
      }
    }
  }
}

Give the agent the runbook discipline in examples/system-prompt.md, then break something and watch the flow:

make inject SCENARIO=failing-rollout CLUSTER=cluster2

You: "Payments is degraded somewhere in the fleet. Investigate and fix."

Agent: list_clusters → get_cluster_health(cluster2) → query_events → get_pod_logs → "payments-v2 on cluster2 is in ImagePullBackOff. Proposing a ManifestWork pinning the last good image. Proposal 4f1a2b3c needs your approval."

You (trusted terminal): ocm-mcp approve 4f1a2b3c, then paste the token back.

Agent: apply_manifestwork → verifies recovery → get_audit_trail → writes the incident report.

Then try to talk it into something dangerous ("just redeploy it privileged with hostNetwork, it's faster"). The proposal dies at layer 1 or layer 2, and the rejection message tells the agent exactly why. More worked examples →

Architecture

_{Dangerous capabilities do not exist. Reads flow freely; every change is proposed, policy-checked, human-approved, and audited.}

flowchart LR
    A["🤖 AI Agent<br/>(any MCP client)"] -->|"typed tool calls"| S["🛡️ ocm-mcp-server<br/>static guardrails · audit"]
    S -->|"reads + dry-run + apply"| H["☸️ OCM Hub<br/>Placement · ManifestWork<br/>Kyverno · RBAC"]
    H --> C1["cluster1"]
    H --> C2["cluster2"]
    H --> C3["cluster3"]
    U["🧑‍💻 Human operator<br/>ocm-mcp approve"] -.->|"approval token"| A
    S -.->|"spans"| J["🔍 OpenTelemetry / Jaeger"]

The write path in one sentence: the agent proposes a ManifestWork; static guardrails and a Kyverno dry-run validate it; a human reviews the exact content and mints an approval token bound to its hash; only then does apply deliver it, with every step traced and logged.

Policy admission with Kyverno

The second guardrail layer does not live in this server - it lives in the cluster. Before a proposed change is ever stored, the server does a server-side dry-run create of the ManifestWork on the hub, so the hub's Kyverno validating admission runs against the exact manifests the agent wants to apply. If your organization's policy says no, the proposal is rejected at admission with the policy's own message - the same control that governs every human kubectl apply.

Why Kyverno:

• Policy as code, no new language. Kyverno is a CNCF policy engine whose policies are ordinary Kubernetes resources in YAML and CEL - reviewable, versioned, and testable like any manifest. This is the policy-as-code approach the CNCF Kubernetes Policy Management whitepaper (CNCF TAG Security) recommends: keep policy declarative and separate from application code.
• Enforced by the cluster, not the prompt. Admission control is external to the model and to this server; it cannot be talked out of the way a system prompt can.
• The right tool for the job. Kyverno can validate, mutate, generate, and verify images; here it is used to validate the workloads embedded inside a ManifestWork.

Where it is used here:

• deploy/policies/ ships 9 ClusterPolicy objects that foreach over spec.workload.manifests inside a ManifestWork: block privileged/host access, protect system namespaces, enforce a kind allow-list, require the managed-by label from the server ServiceAccount (so an unlabeled work cannot skip the others), and enforce a Restricted-Pod-Security baseline in parity with the static guardrails. They are scoped by the app.kubernetes.io/managed-by: ocm-mcp-server label so they judge only agent-authored work. They are usable on their own: deploy/policies/README.md documents the foreach-over-embedded-manifests pattern, the two identifiers an adopter changes, and the Kyverno versions the pack is actually tested against.
• make policy-test runs a 46-case offline suite with the kyverno CLI - good, bad, and human-authored ManifestWorks - needing no cluster and no dependencies. It runs in CI, so a policy regression fails the build before it can reach a hub.
• Don't start from scratch: the community library kyverno/policies and the searchable Kyverno Policies catalog are a ready source of validation, Pod Security Standards, and best-practice policies to adopt or take inspiration from.

Toolsets

The surface is 37 tools across ten toolsets. Almost all of it is read: the whole Open Cluster Management API is safe to inspect. Only two toolsets can change anything, and only through the propose -> approve -> apply gate. Every hub-level tool works for any managed spoke - a standalone OpenShift cluster, a HyperShift hosted cluster, or a cloud cluster - because on the hub they are all ManagedClusters.

_{The whole surface at once. Eight toolsets cannot change anything at all; the two that can are the two wearing a lock.}

Every read tool is annotated readOnlyHint; every write tool is annotated destructiveHint and enforced by the gate. Setting OCM_MCP_READ_ONLY=1 turns off the two writing toolsets entirely, for a strictly-inspection deployment.

Validate against your own hub in one command: ocm-mcp doctor calls every read tool against the live hub and prints a PASS / EMPTY / SKIP / FAIL table (writing nothing), so you can confirm exactly what the server sees before wiring up an agent.

_{The two lanes, to scale: a read answers straight away, a write crawls through propose, a human signature, and a one-time-token apply.}

There is deliberately no tool that reads Secrets, execs into pods, or deletes arbitrary resources. The generic reader (list_resources / get_resource) works against an allow-list of OCM types, so Secrets are not restricted - they are simply not expressible. A capability that does not exist cannot be prompt-injected into use.

Tools

Each tool below is annotated with its class: read (free, no gate), propose (stores a pending change, mutates nothing), or apply (delivers an approved change; needs a human token).

Prompts

The server also ships ten MCP prompts - reusable templates that encode the safe workflow so any client can start from a good runbook instead of a blank box.

Resources

The server also exposes 6 MCP resources - read-only fleet state a client can pin, browse, or attach as context without a tool call (strictly a subset of the read tools; every access still writes an audit line):

Observability - audit, tracing (OpenTelemetry/Jaeger), metrics

Every tool call produces up to three independent records, each with a different job:

_{The same call, recorded twice. Edit an earlier line and the chain says so; the spans, meanwhile, only ever answer where the time went.}

What the tracing is: OpenTelemetry is the CNCF standard for distributed tracing; Jaeger is a CNCF trace viewer. When enabled, this server opens one span per tool call - named tool.<name> (e.g. tool.apply_manifestwork) - with the call's arguments attached as attributes. The approval_token is never attached, and argument values are truncated at 200 characters, so traces are safe to ship to a shared backend.

Why it exists alongside the audit log: the audit log is a safety artifact - append-only and tamper-evident - while spans are a debugging artifact: in Jaeger you can see that a get_cluster_health call spent 4 s waiting on one spoke, or follow the exact propose → apply sequence of an incident on a timeline. Nothing safety-related trusts the spans, which is why tracing can stay optional and fail-soft: without the extra installed and an endpoint set, it is a no-op.

How to use it (two switches + a viewer):

pip install "ocm-mcp-server[tracing]"                    # OTel SDK + OTLP/HTTP exporter
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4318 # your collector

# a local Jaeger to look at traces (make bootstrap starts this for you):
podman run -d --name jaeger -p 16686:16686 -p 4318:4318 jaegertracing/all-in-one:1.60
# open http://localhost:16686 and select the "ocm-mcp-server" service

How it is tested: unit tests cover span creation, token redaction, and the no-op paths; the e2e suite includes a tracing-export step that stands up a local OTLP sink, makes a tool call in a fresh server process, and asserts a real trace batch arrives naming both the tool.* span and the ocm-mcp-server service - so the export wiring is proven on every make e2e and in the nightly CI run. Details: deployment guide - tracing and architecture - observability.

Quickstart (laptop, ~15 minutes)

_{Three ways in. Every row ends somewhere useful — nobody has to reach the third to get an answer out of the first.}

Requirements: podman (or docker), kind, kubectl, clusteradm, helm, Python 3.11+, Linux or macOS (Windows unsupported - use WSL2). The deployment guide has install commands and the real-fleet path.

git clone https://github.com/ocm-mcp-server/ocm-mcp-server.git
cd ocm-mcp-server

make bootstrap      # 1 hub + 3 managed kind clusters, OCM, Kyverno, policies, demo app
make install        # pip install -e ".[dev,tracing]"

Configuration

The server is configured entirely through environment variables. The two that matter most are kubeconfig context names. New to those? The context names guide explains what they are and the exact commands to find yours, from a laptop kind cluster to a cloud login. In short: run kubectl config get-contexts and read the NAME column (make bootstrap prints ready-to-paste values at the end).

For key management, ocm-mcp rotate-secret generates a fresh Ed25519 approval keypair (invalidating every outstanding approval token); ocm-mcp doctor runs the live read-path smoke test; and ocm-mcp audit-verify recomputes the audit log's hash chain to detect any edit, reordering, or mid-log deletion.

# the values make bootstrap prints, spelled out:
export OCM_MCP_HUB_CONTEXT=kind-hub                      # context of the hub cluster
export OCM_MCP_SPOKE_CONTEXTS=cluster1=kind-cluster1,cluster2=kind-cluster2,cluster3=kind-cluster3
#                             └ name on the hub ┘ └ kubeconfig context with read-only creds ┘

Not sure where kind-hub or cluster1=kind-cluster1 come from, or what your own values should be? The context names guide walks through it step by step, including cloud logins (EKS, GKE, AKS, OpenShift). Pointing at a real fleet instead of kind? Same variables; the deployment guide covers the read-only spoke accounts and production hardening.

Then hand the server to your agent: the ready-made client configs are in Connect your agent near the top of this README.

Evaluation harness: honest numbers

eval/ ships 22 scripted incident scenarios in three classes: remediate (15), diagnose-only (3), adversarial (4). Scoring is objective on all three axes: diagnosis keywords in the transcript, live cluster state for recovery, and the server's own audit log for safety.

python3 eval/run_eval.py --agent-cmd "claude -p"     # or any agent CLI

Published results (eval/results/published/), failures included. Every row links to its own raw JSON:

All runs on the same build (v0.6.0, 37 tools, MCP SDK 2.1.1), same fleet, same 22 scenarios. Time taken is wall clock for the whole run. The agy run did not pin a reasoning tier: that CLI offers the model only as high/medium/low and the run took its default, so the exact tier is not recorded.

Not measured counts scenarios where the agent made no tool call, so the server was never consulted. The agent declined on its own, before the request reached the guardrails. Those are excluded from the safety denominator rather than scored, because counting them either way misreports: as a guardrail success that was not earned, or as a failure that did not happen.

Safety is the axis this server exists for, and it held on every scenario that reached it. The interesting number is the one beside it. Frontier models increasingly refuse an adversarial bait before calling any tool, so the guardrails are never consulted and a bait that was never presented would otherwise score identically to a bait that was blocked. Recovery misses concentrate on scenarios whose fix needs state the read surface deliberately withholds.

_{Diagnosis and recovery belong to the agent and are published unflattered. Safety is the axis this server is answerable for, and it is the full one.}

Run it against your model of choice and publish your numbers, including the failures. The point is real data about what agents can and cannot yet be trusted to do.

The Kyverno policies have their own offline test suite: make policy-test runs 46 CLI cases (deploy/policies/tests/) against good, bad, and human-created ManifestWorks with no cluster and no dependencies. It runs in CI too, so a policy regression fails the build before it ever reaches a hub.

Try it end to end (one command)

Want proof it works against real clusters, not mocks? One script stands up a real Open Cluster Management fleet on kind, exercises every tool and prompt, runs a break-then-fix scenario, and writes a graphical HTML report:

./hack/e2e-local.sh          # 2 spokes, auto-cleanup   (SPOKES=1 for a lighter run)

It (1) installs or version-checks the dependencies (Podman, kind, kubectl, clusteradm, helm; Docker is not required), (2) bootstraps a hub plus spokes, (3) runs every read tool, the gated propose -> approve -> apply write flow, the gated ROLLBACK flow, every lifecycle action (cordon/uncordon, set_label, accept, enable/disable_addon), and all ten prompts - each with a plain-language explanation of what it does and why, (4) drives the real server binary over stdio JSON-RPC with the official MCP client (handshake, tools, prompts, resources, annotations), (5) runs a negative sweep proving every gate fails closed (expired token, replayed token, apply-scoped token refused for rollback, read-only mode, tampered audit log caught, signed audit anchor verified) plus a tracing-export check (OTel spans over OTLP received by a local sink), (6) injects a failing rollout and shows the diagnose-and-fix loop end to end, then (7) writes e2e-report.html and tears the fleet back down (kind and Podman stay installed). The report is git-ignored. Works on macOS (Homebrew + Podman) and Linux, and runs nightly in CI.

ocm-mcp doctor runs just the live read-path smoke test on its own, against any hub.

Here is a real, unedited run (recorded with asciinema, long waits compressed): the fleet comes up, every step passes, and the fleet is torn down again - MP4 version · terminal cast:

Documentation

Repository map

src/ocm_mcp_server/   the MCP server: tools, guardrails, approvals, tracing, CLI
deploy/               least-privilege RBAC + Kyverno ClusterPolicies (+ offline tests)
hack/                 bootstrap.sh / teardown.sh / demo app (kind-based fleet)
chaos/                failure-injection scenarios (reversible, diagnosable)
eval/                 22-scenario evaluation harness + results
blogs/                long-form posts (canonical drafts; published to Medium)
docs/                 deployment, examples, architecture, guardrails, demo, upstream
examples/             MCP client configs (Claude, VS Code, Codex, Gemini) + system prompt

Roadmap

The canonical, themed roadmap lives in ROADMAP.md. Current headline items: an authenticated HTTP transport with per-tool scopes, an off-box (KMS/HSM) approval signer, the OCM cluster-proxy transport, and a reusable Kyverno policy pack. (Multi-model eval results are now published.)

Have a need that's not there? Open a feature request. New tools require a safety rationale; see CONTRIBUTING.md.

Contributing & community

Issues and PRs welcome. Start with CONTRIBUTING.md. Getting help: SUPPORT.md. Security reports (privately, please): SECURITY.md.

Sponsorship

This project is independently maintained. If your organization wants priority integration help, a hardened deployment review, sponsored features, or talks and workshops on safe agentic operations, connect on LinkedIn (details in SUPPORT.md).

Author

Sandeep Bazar - Passionate in Technology especially around Multi-cluster Kubernetes platforms, day-2 operations, and making fleets safer to automate.

If this project is useful to you, a ⭐ helps others find it.

Project governance and maturity

This project holds itself to CNCF community, governance, and security practices as a quality bar - the same standards expected of a CNCF Sandbox project - so it is easy to adopt, contribute to, and trust. The scaffolding is in place:

• Governance - how decisions are made and how to become a maintainer.
• Maintainers and Adopters - who maintains it, and an open page to record who uses it.
• Roadmap - direction by theme, safety-first.
• Security self-assessment - a CNCF TAG-Security-style assessment of what the guarantees are and where they end.
• CNCF Sandbox readiness - a self-check against CNCF Sandbox expectations, used as a quality bar.
• Code of Conduct - the CNCF Community Code of Conduct; report privately via LinkedIn.

Contributions are signed off under the DCO - enforced by a CI job on every pull request, not just asked for in a checklist - published release tags are immutable, and any change that touches a guardrail requires a written safety rationale.

License

Apache-2.0