convez.ai
MCP ServerDevOps & InfraestruturaAvançado
romaine-life

MCP Tank Operator

romaine-life/mcp-tank-operator

Gerenciar sessões de IA em cluster via MCP — criar, deletar, monitorar e configurar.

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP em cluster que permite gerenciar sessões de IA (Claude, Codex, Antigravity) em nome do proprietário. Oferece controle completo sobre criação, exclusão, monitoramento de transcrições e configuração de sessões, com integração nativa ao orquestrador Tank e suporte a slots de teste Glimmung.

Instalação

$npx convez-ai romaine-life/mcp-tank-operator

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
12%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
7

Casos de uso

  • Automação de criação e limpeza de sessões de chat IA em produção
  • Monitoramento de transcrições e triagem de sessões travadas
  • Validação de imagens de session-runner em slots de teste
  • Integração de SpireLens MCP entre sessões gerenciadas

Veredito

Ferramenta especializada para operadores de infraestrutura Romaine que precisam gerenciar o ciclo de vida de sessões IA em Kubernetes. Requer conhecimento profundo da plataforma Tank e não é adequada para uso genérico fora desse ecossistema.

README

A tradução para o português deste README ainda está na fila.

mcp-tank-operator

In-cluster MCP server that lets a tank-operator session pod manage other sessions on behalf of its owner.

Layout

  • src/ — Python MCP server package.
  • Dockerfile — image build for romainecr.azurecr.io/mcp-tank-operator.
  • chart/ — Helm chart synced by ArgoCD.

Images are SHA-tagged from main; .github/workflows/build.yml pushes the image and commits the matching chart tag.

Tools

  • list_sessions() — sessions owned by the calling user.
  • list_session_refs() — low-noise list of session ids and Tank UI display names.
  • resolve_session(session_ref) — resolve a Tank UI display name or session id to the full session record.
  • get_session_capability_context(capability, session_id) — warm documentation for rare Tank session capabilities. Currently covers spirelens_mcp, including the native MCP endpoint, direct auth.romaine.life SSH certificate path, and the difference from Glimmung run callback URLs.
  • verify_spirelens_session_access(session_id) — read-only inspection for a caller-owned session's SpireLens MCP wiring: selected capability, /workspace/.mcp.json server entry, local proxy target, and expected host lifecycle tools.
  • read_transcript(session_id, ...) — read a caller-owned session's conversation transcript (projected rows + pagination cursors). Reads the durable Postgres projection, so it works even after the target session's pod is gone — useful for triaging a stuck sibling session before deciding to prompt, delete, or escalate.
  • get_session_run_options() — read Tank-owned create modes, SDK chat modes, provider model/effort lists, retired create modes, and defaults. Use this before choosing a non-default mode, model, or effort; Tank remains the validator and returns actionable errors for unsupported values.
  • create_session(mode) — spawn a new session pod. Current chat modes are claude_gui and codex_gui; default is claude_gui.
  • delete_session(session_id) — delete one of the caller's sessions.
  • set_session_name(session_id, name) — set or clear the friendly display name.
  • set_test_environment(session_id, ...) — update the Tank UI test workflow link state.
  • set_pull_request_link(session_id, url) — update the Tank UI PR link for the active test workflow.
  • get_session_url(session_id) — tank UI URL for an existing session; accepts either an id or display name.
  • send_prompt(session_id, prompt, ...) — fire-and-forget follow-up prompt to an SDK chat session.
  • spawn_run_session(prompt, mode, ...) — create a fresh SDK chat session, wait for its pod to become ready, then queue the first prompt.
  • spawn_test_slot_session(slot_name, prompt, mode, ...) — create a fresh SDK chat session through a Glimmung test slot's own Tank orchestrator, then queue the first prompt. Use this for test-slot validation; it requires a slot name such as tank-operator-slot-2 and refuses production-ish targets.
  • point_slot_session_image(slot, codex_image, claude_image, antigravity_image, git_ref) — point a Glimmung test slot at a branch-built session image so NEWLY-created sessions in that slot boot it (the same image lever production uses, no runtime overlay). Covers all three session-runner providers (claude / codex / antigravity); set one or several at once. For antigravity specifically this is the supported validation path, since the Go antigravity-runner has no running-pod hot-swap via apply_test_slot_hot_swap. Complements Glimmung's apply_test_slot_hot_swap, which only patches already-running pods. The image must already exist in ACR (build it via the tank-operator session-images-build.yml workflow first); the production scope is refused server-side, so this can only repoint test slots. Targets the slot's own orchestrator (tank-operator.<slot>.svc), where the test-env gate is on.
  • get_slot_session_image(slot) — report what session image NEW sessions in a test slot will boot (the current override, or override_set: false). Read-only.
  • clear_slot_session_image(slot) — clear a slot's session-image override; new sessions revert to the chart-pinned image.

Auth

Inbound: kube-rbac-proxy validates the calling session pod's projected SA token via TokenReview + SubjectAccessReview against the synthetic mcp.tank-operator.io/servers/tank-operator resource. This gates whether the pod can reach this MCP server at all; per-call identity is a separate layer below.

Outbound: every tool authenticates to the orchestrator with the calling pod's auth.romaine.life service-principal JWT, forwarded by the pod's mcp-auth-proxy sidecar in the X-Auth-Romaine-Token header. mcp-auth-proxy exchanges /var/run/secrets/auth.romaine.life/token at auth.romaine.life/api/auth/exchange/k8s for a role=service JWT and forwards it; this server extracts it into the SERVICE_BEARER ContextVar and passes it as Authorization: Bearer on the outbound /api/internal/sessions/* call. The orchestrator verifies the JWT and treats its actor_email claim as the owner — the MCP server never sees or accepts an owner email.

The pre-#486 IP-tail identity path (X-Forwarded-For → caller_pod_ip query param) was retired in Stage 4. See romaine-life/tank-operator#486.

Ver o original no GitHub

Bom para

Relacionados