The week on the bench · Monday to Sunday, UTC

Week 25, 2026 · 15 to 21 June

8 entries published: 3 tools, 1 build, 2 guides, 2 news. This page lists what was published; it is not a newsletter issue.

The week as a scope trace

One channel per kind, one pulse per entry at the hour it was published. Taller pulses have more points.

Build of the week · BB-0092

Saras 2.0: a Raspberry Pi 4 rover that runs Claude Code as its brain

Tech With Anant rebuilt the Saras rover around Claude Code running on a Raspberry Pi 4. A small MCP server exposes the motors, encoders and camera as tools, and a CLAUDE.md file sets the robot's identity. An LD06 lidar handles obstacle detection. The 24-minute video shows the rover in a live test, driving toward a named object around obstacles.

Shown workingBy techwithanant-yt · youtube.com · Fri 19 Jun, 12:00 UTC

Tools3

All tools
  1. 4 mo ago

    ToolBB-0091github.com0 starsPython

    servomotor-mcp: Claude Code drives Gearotons M17 RS-485 servomotors

    Gearotons' servomotor-mcp drives its M17 closed-loop NEMA-17 servomotors over RS-485 from Claude Code. It auto-detects motors on the bus, wraps each of the 48 firmware commands as a tool and adds high-level moves; a mock backend runs without hardware. A terminal recording and a 50-second video show Claude Code turning a bench M17 and reading its encoder.

    Shown working
  2. 4 mo ago

    ToolBB-0088github.com7 starsPython

    vibe-hardware: agent skills from idea to firmware, PCB and enclosure

    vibe-hardware installs skills into Claude Code or Codex that take an idea to ESP32 firmware, a KiCad board, a build123d enclosure and a JLCPCB order. Its example is a desk pager on an M5StickC S3 that buzzes when a Claude Code session needs input. The README says the firmware works; the PCB and shell are still stubs.

Builds1

All builds
  1. 4 mo ago

    BuildBB-0092youtube.comBuild of the week

    Saras 2.0: a Raspberry Pi 4 rover that runs Claude Code as its brain

    Tech With Anant rebuilt the Saras rover around Claude Code running on a Raspberry Pi 4. A small MCP server exposes the motors, encoders and camera as tools, and a CLAUDE.md file sets the robot's identity. An LD06 lidar handles obstacle detection. The 24-minute video shows the rover in a live test, driving toward a named object around obstacles.

Guides2

All guides
  1. 4 mo ago

    GuideBB-0089iglos.com

    Linutronix on the lab behind an agent that reserves, flashes and power-cycles real boards

    Linutronix describes an off-the-shelf AI agent that tests embedded boards in its lab through a small MCP server. In the example session the agent reserves a board, flashes it, reads the console and power-cycles it after a bad flash. The post argues that the safety comes from the lab underneath. It shows one screenshot and no logs.

  2. 4 mo ago

    GuideBB-0086developer.espressif.com

    Espressif's notes on building a Zephyr BLE and SoftAP provisioning app with Cursor

    Ricardo F. Tafas Jr of Espressif describes building Zephyr provisioning over BLE and SoftAP for the M5Stack ESP DualKey with Cursor. He tested each step on the bench through build, flash and UART traces. His practices include specs first, proving a module in a sample app, a journal of failed fixes and no edits to upstream Zephyr.

    Claimed
  1. 4 mo ago

    NewsBB-0090anthropic.com

    Project Fetch phase two: Claude Opus 4.7 in Claude Code writes the sensor code for a robot dog

    Anthropic repeated its robot-dog experiment with Claude Opus 4.7 in Claude Code, a researcher approving each command. On every task a 2025 human team had completed, including connecting to the camera and lidar, Claude was at least ten times faster. Anthropic says its attempts to knock the beach ball back were poorly controlled, and it never finished the autonomous fetch.

  2. 4 mo ago

    NewsBB-0085arxiv.orgCurator pick

    Embedded Arena: agents deploy models to microcontrollers with hardware feedback

    Zhihan Zhang and colleagues, mostly at the University of Washington, published Embedded Arena, a benchmark that scores agents' model and firmware edits on microcontrollers. Boards include a MAX78000, an ESP32-S3 and an STM32N6. The paper reports Claude Opus 4.7 and Gemini 3.1 Pro failing every deployment without hardware feedback; with it, the first success came within 3 iterations.

    Claimed