LabWired

Agents call LabWired, and LabWired does not act in other systems on their behalf. Its tools work on its own simulator, part catalog and labs, and the hosted builder compiles PlatformIO projects before a run. The GitHub Action posts its results back inside the customer's own workflow, and flashing a board on the desk runs through the customer's local tools.

LabWired offers no connectors for an agent to take action in outside services such as a ticketing system or a repository host.

Sourcedocs.labwired.com/agent/toolsread 2026-10-08

LabWired's own CLI agent runs a loop of write, check, fix and green: it writes firmware, runs it on the twin or a connected board, reads the result and patches the code until the checks pass. Every step is headless, so the loop can run with nobody present, and LabWired says an agent can run fifty loops in parallel across boards and configurations.

Multi board and multi MCU systems can sit on one canvas and run together. Beyond the agent's loop and those parallel runs, LabWired offers no workflow designer and no way for several agents to hand work to each other.

Sourcelabwired.com/agent.htmlread 2026-10-08

The knowledge on offer is LabWired's own part catalog. Tools look up parts by manufacturer part number or alias, describe boards, pins and buses, and return datasheet text: LabWired extracts each datasheet page by page, the agent searches that text and reads the best matching pages in full, and answers cite the page and a hash of the document so the reference holds later.

A schematic can be imported into a twin for one project. LabWired does not offer a searchable store of the customer's own documents.

Sourcedocs.labwired.com/agent/toolsread 2026-10-08

The CLI agent runs its write, check and fix loop with no approval step, and LabWired describes the loop as able to run with no person present. The MCP tools run when called. After a run, the agent opens a Playground link so a person can look at the board, and the verify tool reports whether the result passed, but nothing pauses the agent for a person to approve an edit or a flash beforehand.

Sourcelabwired.com/agent.htmlread 2026-10-08

Single sign on comes with the Enterprise plan, along with team seats and an audit log, though LabWired says nothing further about roles or permissions. Agents sign in to the hosted MCP server through the browser with the customer's LabWired account, using OAuth or device code tokens.

LabWired names no certification, attestation or trust center. Its privacy notice covers TLS for every connection, daily compressed backups with a seven day restore window and a 60 day backstop, deletion on request, redaction of hosted agent logs and a statement that customer firmware, prompts and twin runs are not used to train models by default.

Sourcelabwired.com/pricing.htmlread 2026-10-08

An audit log is listed on the Enterprise plan, with nothing said about what it records or how long it keeps it. Every run writes result.json, a VCD trace, UART and RTT logs, display frames and a cycle by cycle program counter history, and the VS Code extension adds a register inspector and a timeline view. Those describe what the firmware did.

Hosted agent conversations, tool calls included, are stored, though LabWired does not describe a customer view of them.

Sourcelabwired.com/pricing.htmlread 2026-10-08

Runs leave state behind that can be picked up later. Each run returns a snapshot ID that the inspect tool reads back for deeper CPU and peripheral state, Pro keeps validated private twins, and labs persist behind share links. Hosted agent conversations, including prompts, outputs and tool calls, are stored as well, while local agent sessions stay on the developer's machine.

These hold boards, runs and transcripts. LabWired keeps no store of facts or conventions that the agent consults at the start of its next task.

Sourcedocs.labwired.com/agent/toolsread 2026-10-08

Enterprise customers can self host LabWired or run it air gapped, so firmware never leaves their network, with custom device onboarding and co-simulation included. The open source engine, the local MCP server and the Docker image also run on the customer's own machines or CI runners without a cloud account or license server. The hosted service runs in the United States and other regions where its providers operate, with no region choice offered.

Sourcelabwired.com/pricing.htmlread 2026-10-08

Board configs come ready for STM32 parts from the F103 to the H735 and WBA52, Nordic nRF52832, nRF52840, nRF5340 and nRF54L15 development kits, the Raspberry Pi Pico and ESP32, ESP32-C3 and ESP32-S3 devkits. A parts library adds sensors such as the BME280 and MPU6050, SSD1306, Nokia 5110 and e-paper displays, a thermal camera, motor drivers, test instruments and IO-Link and UDS stacks.

Public labs, a CI starter template, integration templates and part packs round it out, and a twin pack can be imported from a schematic. These are boards and labs to build on, and the LabWired Agent is the one agent offered.

Sourcedocs.labwired.comread 2026-10-08

Work starts when someone runs labwired agent at a terminal or opens the LabWired Agent web page, or when an outside coding agent calls the MCP tools. In CI, the labwired/firmware-test action or the ghcr.io/w1ne/labwired Docker image runs when the customer's own GitHub Actions, GitLab CI or self hosted runner fires on a change, and runs can be pinned to a git commit.

LabWired itself has no scheduler, webhook or event listener that starts work on its own.

Sourcelabwired.com/ci.htmlread 2026-10-08

Customers choose what powers the CLI agent: a local model, their own Codex or Claude traffic billed separately by that provider, or LabWired's hosted inference, which on Pro includes a LabWired model for Zephyr work. Enterprise adds an on premises LabWired model with private weights. Outside coding agents that connect over MCP keep whatever model they already use.

Sourcelabwired.com/agent.htmlread 2026-10-08

Coding agents connect to a hosted MCP server at api.labwired.com/mcp and sign in through the browser with their LabWired account, and a one line installer registers it with Claude Code, Codex, Cursor, VS Code or OpenCode. The tool reference lists every tool, from search, list, describe and validate to compile, run, inspect, verify and lab, plus part, datasheet, catalog, export and device validation tools. They compile firmware, run it, return snapshots and create shareable labs on LabWired itself, through the same API the Playground uses.

A local stdio server from npm, the labwired CLI and a Python SDK that drives the engine from pytest round out the developer surface. The engine accepts any standard ELF, whether it was built with PlatformIO, make or west.

Sourcedocs.labwired.com/agent/toolsread 2026-10-08

Firmware changes get a real verdict. A YAML test script asserts with uart_contains, uart_regex, rtt_contains and memory_value under step, cycle, wall time, output size and stall limits, can end early once its assertions pass, and returns pass, fail or error in result.json, exit codes and JUnit XML. Runs are deterministic, so the same binary and inputs give the same result on every machine.

For agent runs, the verify tool is the pass or fail check, and the labwired/firmware-test action runs the same tests on each change, in under two minutes on free GitHub runners by LabWired's account. A validation matrix per chip and peripheral is refreshed nightly, and LabWired compares its peripheral models with physical silicon on a hardware bench. The most complete models today cover UART, GPIO, interrupts, timers, clocks and flash, with I2C, SPI, ADC and DMA growing and USB, CAN and Ethernet partial.

Sourcedocs.labwired.com/ci_test_runnerread 2026-10-08

Agents work with LabWired through MCP tool calls and the CLI, reading structured results rather than screenshots. The Playground is a browser app for people to view boards and share labs, and no browser or desktop is operated by the agent on software LabWired does not control.

Sourcedocs.labwired.com/agent/mcpread 2026-10-08