Autonomous agents.
Real cloud operations.

The AI tooling market split early into two camps. The first — Copilot, Cursor, Claude Code, Continue — lives in your IDE and helps you write code. The second — n8n, Zapier AI, custom LangGraph deployments — lets you wire LLMs into business workflows, but assumes someone else built the infrastructure layer. Neither camp handles the actual work of operating a cloud.

The reason is unglamorous: cloud ops requires holding production credentials, recording every action, enforcing approval on destructive operations, capping spend per agent, and translating one intent ("deploy this") into five different cloud APIs. None of that is interesting research; all of it is required before a sane company will let an LLM near its AWS console. vxcloud is the platform layer that ships all of it as a default — so you can spend time on the agent prompts and the use cases, not on rebuilding RBAC for the fourth time.

AgentControl is the training plane: a managed pipeline for fine-tuning small LLMs against your internal data and serving them as bound agents. The classic shape is: import a Git repo as a dataset, point at a Claude Haiku or Llama-3 base model, train, and the resulting agent is wired into a marketplace slot that anyone in the workspace can call.

import vxsdk
c = vxsdk.Client.load_from_vxcli()

# Import the repo as a JSONL dataset
ds = c.agentcontrol.github.import_dataset(
    repo="acme/platform", branch="main",
    include=["**/*.py", "docs/**/*.md"],
    exclude=["tests/**", "vendor/**"],
    dataset_name="acme-platform-snapshot",
)

# Fine-tune Claude Haiku on it
job = c.agentcontrol.fine_tuning.create(
    base_model="claude-haiku-4-5",
    dataset_id=ds["id"],
    name="acme-platform-coder-v1",
    hyperparameters={"epochs": 3, "learning_rate": 1e-5},
)

# Block until terminal status (succeeded / failed / cancelled)
job.wait_for_completion(timeout_seconds=3600)
print("model_id:", job.data.get("model_id"))

# The new agent is now in the marketplace under the trained model name
c.agentcontrol.agents.execute(
    agent_id="acme-platform-coder-v1",
    input="Refactor the metrics module to use OpenTelemetry SDK directly",
)

• Per-workspace HashiCorp Vault — agents never see plaintext cloud credentials.
• RBAC scoping — an agent gets exactly the toolset you grant it, no implicit root.
• Audit trail — every agent action recorded with input, output, and the human who delegated it.
• Approval gates — `resolveApproval()` pauses on destructive operations until a human signs off.
• Budget caps — per-agent daily and per-action spend ceilings prevent runaway loops.
• Bring Your Own Model — Anthropic, OpenAI, Gemini, Deepseek, Llama, Mistral, Ollama all supported via the chat envelope.
• Replay — re-run any past agent session deterministically for debugging or compliance review.
• Cross-cloud — agents work the same against AWS, GCP, Azure, Alibaba, and Linode — no per-cloud agent rewrites.