Skip to content

Annealage docs

Annealage brings agent-driven tools across hardware development, from schematic design and review to firmware, embedded interfaces, hardware-in-the-loop testing and printable parts. Some agents run locally, others are hosted, and MCP connects agents to the work where it happens.

These docs cover Canvas (in private use), Pod for hardware-in-the-loop testing, and Mesh for 3D-printable parts.

Pod and Mesh are available now. Canvas is in private use; its guide and setup instructions are for authorised users.

  • Set up Pod: start here if you verify firmware on real hardware. Pod is network-attached, so the agent flashes and tests a board over Wi-Fi without a wired host.
  • Run Mesh: start here if you design enclosures, brackets or other printed parts. Mesh runs locally with its own agent chat pane beside the 3D viewer.

Each is independent. Pick the one that matches what you are building; you do not need the others to use it.

A Wi-Fi hardware-in-the-loop test rig on an off-the-shelf Raspberry Pi Pico 2 W (RP2350). The debugger runs on the pod itself: PIO SWD, an ADIv5 stack, a Cortex-M layer, a flash loader, and a GDB server, so it flashes, debugs, resets, and logic-analyses a device-under-test over the network with no wired host and no separate probe. A stock arm-none-eabi-gdb connects through the pod; USB/IP attaches the DUT’s USB to your host. Dev boards with their own programmer use the same verbs with no pod, through pyOCD on the host.

An agentic CAD workbench for 3D-printable parts. A live 3D viewer sits beside an agent chat pane; the agent writes CadQuery geometry and the regenerated STL appears within a fraction of a second. Click the model to pin feedback to exact coordinates, and the agent pins its callouts back.

Terminal window
uvx annealage-mesh ./part

An in-process LVGL development harness for the MicroPython unix port. It serves a /reload HTTP framebuffer viewer with browser-driven virtual touch, so you see the GUI in a browser while an agent screenshots it, taps widgets, and inspects the widget tree over MCP. The same code path runs standalone (SDL plus the web viewer) or embedded, serving a real MCU’s framebuffer over HTTP.

Terminal window
pip install annealage-canvas

Canvas and Pod each run an MCP server; register it with any MCP client, such as Claude Code. Canvas serves MCP over HTTP from the running harness:

Terminal window
claude mcp add annealage-canvas --transport http http://localhost:5050/mcp

Pod’s server is the pod-mcp console script over stdio:

Terminal window
claude mcp add pod -- pod-mcp

Mesh brings its own agent in the chat pane (Claude, Codex or omp). An agent running elsewhere works through two JSON files in the served folder plus a few MCP tools; see Mesh.

Connect your agent has the full setup.

Loom and Trace are not publicly available, and neither is documented here.