01
Plan
Gather and propose
Gather relevant state and propose the exact read and write operations required.
Open source · Apache-2.0
AI agent for Kubernetes. Approval-gated by design.
Skyflo is an open-source Kubernetes AI agent for troubleshooting and approved operations. It collects cluster context, proposes a plan, waits for approval before mutations, executes through typed tools, and verifies the resulting state.
Self-hosted
Typed MCP tools
Operator-gated writes
Apache-2.0 and self-hosted. No Skyflo telemetry. Model requests go only to the provider you configure.
Illustrative UI · not a production transcript
cluster connectedOP
The payments deployment is crash-looping. Find the cause and prepare a safe fix.
Evidence gatheredread-only
Pod state
Recent events
Deployment spec
Proposed operation
kubernetes.patch_resource
- target
- deployment/payments
- namespace
- production
- change
- memory limit · 256Mi → 512Mi
Nothing mutates until an operator approves
Illustrative UI, not a transcript or a measured production incident.
A separate open-source project for Kubernetes operations, distinct from the Skyflo Desktop harness.
Explore Desktop HarnessThe control loop
Read freely. Write deliberately.
The agent can gather evidence without slowing the operator down. The moment a tool would change live state, the engine stops the run and asks for explicit approval.
02
Approve
Stop before the write
Hold each mutation in an approval state containing the typed operation and arguments.
03
Execute
Issue only the approved operation
Run the approved operation through the declared typed-tool boundary.
04
Verify
Check the result
Read the state again and compare it with the original intent.
Reads can auto-continueMutations always pauseResults are checked against intent
Operations surface
One agent. Four typed toolsets.
The agent loads declared schemas for the operation at hand. Read paths stay fast; every traffic-affecting or state-changing path inherits the same gate.
22 typed tools
Kubernetes
Discover, inspect, apply, scale, restart, drain, and roll back cluster resources.
Pods & workloadsEvents & logsDry-run apply
16 typed tools
Helm
Inspect releases and values, render templates, then gate installs, upgrades, and rollbacks.
Release statusRendered diffRollback
13 typed tools
Argo Rollouts
Read rollout state and keep pause, resume, promote, cancel, and abort behind approval.
Canary statusPromotion gateAbort
13 typed tools
Jenkins
Read jobs, build history, logs, and SCM context before gated build control.
Build logsSCM contextTrigger & stop
Built for the production boundary
The operator stays in the loop.
Skyflo runs inside your cluster. The model can propose work, but declared tool schemas, engine policy, and cluster RBAC define what can actually happen.
Read the approval modelExecution boundary self-hosted
01
Operator
02
Engine gate
03
MCP tools
The gate lives in the engine
Approval is enforced by runtime tool annotations, not a prompt or a UI preference.
Typed tools, not shell text
MCP schemas bind every operation to declared parameters and structured results.
A record of the whole loop
Requests, plans, approvals, tool calls, and results are persisted for audit and replay.
Run it in your cluster
See the gate before you trust the agent.
Install the Apache-2.0 project, connect a cluster, and start with a read-only investigation. Mutations will wait for you.
$
curl -fsSL https://skyflo.ai/install.sh | bash