Skip to content

1.2 AI ERA

Choose a tag to compare

@Butterski Butterski released this 05 Oct 19:24
· 49 commits to master since this release

HLBuilder can now work with an LLM. Connect your own client over MCP, or turn on the assistant inside the builder. Either way the model can only propose changes: you review each proposal on a preview canvas and apply or reject it.

New Features

MCP server: Connect Claude Code, Cursor, VS Code, LM Studio or any MCP client to /mcp. The client can read your builds, validate the network, generate configs, search the hardware and service catalogs, and propose changes. Works on self-hosted instances and on the hosted one. Setup guide: docs/MCP.md

Personal access tokens: Create tokens in Settings, read-only or read + propose, optionally limited to a single build, with an expiry. Tokens are stored as SHA-256 hashes and shown once. A token is always required, including on instances that run without login.

Proposals instead of edits: An LLM never writes to a build. It sends a change set that is checked with a dry run (including the IP addresses devices would get). You see it on a read-only preview with new, changed and removed parts marked, then choose Apply or Reject. Apply keeps edits you made in the meantime, and one Ctrl+Z undoes it.

In-app assistant: An optional chat panel in the builder, off by default. Bring your own provider and key: Anthropic, OpenAI, Google Gemini, OpenRouter, Ollama, or any OpenAI-compatible endpoint. It uses the same tools and the same proposal flow as MCP clients.

Key protection you can check: Provider keys are encrypted with AES-256-GCM, tied to your account, and never sent back to the browser. Settings explains why the key is kept on the server, shows the stored record, states what it does not protect against, and links to the code. Details: docs/AI-ASSISTANT-SECURITY.md

Settings page: New page for appearance, the AI assistant and MCP access. Theme settings moved here from Profile.

Also new since 1.1

Virtual networks: Open a host's virtual network to wire its VMs and containers to virtual switches and the physical uplink. See docs/VIRTUAL-NETWORKS.md

Custom services: Create your own services in the service catalog.

Fixes

  • Reassign IPs no longer fails until the page is refreshed.
  • Firewall protection zones are traversed correctly.
  • Deleting a build that contains devices no longer fails.
  • A device's power draw is saved and restored.

Upgrading

  • New database tables are created automatically on start.
  • Public instances (login enabled, release mode): set SECRETS_KEY (openssl rand -base64 32). Without it the assistant stays switched off.
  • Behind your own reverse proxy: forward /mcp and /api/assistant/ unbuffered, as frontend/nginx.conf does.
  • Building from source now needs Go 1.25.
  • New optional variables: MCP_ENABLED, ASSISTANT_ENABLED, MCP_ALLOWED_ORIGINS, PUBLIC_APP_URL, ASSISTANT_ALLOW_PRIVATE_ENDPOINTS. See .env.example.

Known limits

  • Claude Desktop connects through the mcp-remote bridge (snippet in the setup guide).
  • The claude.ai and ChatGPT web connectors require OAuth and cannot use a token.
  • Provider integrations are new. If one misbehaves, please open an issue with the provider and model name.