Plugin catalogue · Installation
The agentic-engineering desired-state manifest
is a provider-neutral, copy-paste onboarding document for a new assistant. Open the assistant in the
consumer repository, paste the complete JSON, and ask it to reconcile the desired state. It points to
the latest reviewed plugin for generic role logic and to the consumer's AGENTS.md for organization,
trust, cadence, memory, and maintainer-channel configuration. It also requires the assistant to report
unsupported native capabilities instead of silently weakening the deployment. The manifest carries
separate thin schedule prompts for the Agentic Engineer and the Agent Improver; each resolves its
cadence and deployment facts from the canonical consumer instructions. Spend stewardship has no
schedule of its own. It runs inside the engineer's loop only after explicit maintainer opt-in through
spec.roles["agentic-engineer"].spendStewardshipEnabled and a resolving Spend contract.
Existing installations must complete all four migrations, in order, before their next scheduled
run — the
version 2 checklist
(plugin identity), then the
version 3 checklist (retire the
finops-engineer schedule and adopt the Spend contract section), then the
version 4 checklist
(the automated-ai-engineer → agentic-engineer entrypoint rename), then the
version 5 checklist
(explicit default-off spend enablement). Stopping early would resume
unattended writes with the retired FinOps schedule still armed, or with a schedule pointing at an
entrypoint that no longer resolves.
A plugin may bundle MCP servers as well as skills. The
gitops-kubernetes plugin bundles the Flux MCP server
(flux-operator-mcp)
so its gitops-cluster-debug skill — which Requires flux-operator-mcp — can connect to a live cluster once the server binary and cluster access are configured.
The server is authored once as the plugin's .mcp.json
(mcpServers map). How each tool consumes it differs (per ADR 0001):
- Claude Code, Copilot CLI, and VS Code — the bundled
.mcp.jsonis loaded automatically when the plugin is installed. The server binary must be installed separately and have access to the intended cluster.
You only need to write MCP config by hand if you are not installing this as a plugin — then add the
server to your workspace .vscode/mcp.json (note the key there is servers, not mcpServers):
{
"servers": {
"flux-operator-mcp": { "command": "flux-operator-mcp", "args": ["serve"] }
}
}Every path invokes the same flux-operator-mcp binary, so install it first — e.g.
brew install controlplaneio-fluxcd/tap/flux-operator-mcp or go install github.com/controlplaneio-fluxcd/flux-operator/cmd/mcp@latest (it reads your kubeconfig from
KUBECONFIG / ~/.kube/config). See the
Flux MCP docs for read-only mode and remote
transport.
A plugin may also bundle custom agents (subagents). The
gitops-kubernetes plugin bundles
flux-troubleshooter — a read-only
Flux CD triage agent that traces the GitOps dependency chain (source → Kustomization/HelmRelease →
workloads), reads status conditions and controller logs via the bundled flux-operator-mcp server,
and returns a root-cause diagnosis plus the human-applied fix. It has no apply/reconcile/suspend/
delete tool by design, so it never mutates the cluster.
The agent is authored once as agents/<name>.agent.md (Markdown + YAML frontmatter, with the neutral
name/description/tools/model core). The .agent.md filename is what VS Code and Copilot
discover inside a plugin's agents/ directory; Claude Code is filename-agnostic (it reads any
Markdown in agents/ and takes the agent's name from frontmatter), so one file serves all three
tools. The same goes for the tools: allowlist: MCP tools are listed in both spellings — Claude
Code's mcp__<server>__<tool> and VS Code / Copilot's <server>/<tool> — because each tool
ignores entries it does not recognise (per
ADR 0001):
- Claude Code, Copilot CLI, and VS Code — the bundled
agents/directory is loaded automatically when the plugin is installed; in Claude Code the agent is namespacedgitops-kubernetes:flux-troubleshooter.
As with MCP, hand-placing an agent at .github/agents/<name>.agent.md is only for setups that aren't
installing this as a plugin.
The vibe-coding plugin bundles
vibe-coding-companion — a plain-language
build companion for a non-technical audience (design:
ADR 0003). Same delivery rules. Its guardrail requires
the consuming deployment to author a ## Stack map section in its AGENTS.md (see the
plugin README).
The agentic-engineering plugin bundles three agents —
agentic-engineer (the
autonomous portfolio-engineer actor),
portfolio-surveyor (its read-only
survey subagent), and
agent-improver (a meta-engineer that
improves the engineer itself from measured evidence) — alongside its engineering skills (design:
ADR 0002, consolidation:
ADR 0004, spend stewardship:
ADR 0005). Same delivery rules; the
consuming deployment must define the five contract sections (Portfolio map, Trust gate, Cadence,
Memory, Maintainer channels) in its AGENTS.md — plus Agent definition locations and
Authority model if it enables agent-improver, and Spend contract plus explicit maintainer
opt-in if it wants the engineer to steward spend (see the
plugin README).