PIO Agent

[PIO Agent]

PIO Agent

PIO Agent is the open-source, agent-first hardware execution layer for embedded development, built on the PlatformIO MCP runtime.

Release status

The canonical 3.1.0 runtime is published. The 3.2.0 candidate makes the pio-agent CLI the primary adapter (nothing stays running unless asked), and adds a native Codex panel and public plugin listing metadata. See the changelog, compatibility guide, and distribution readiness for implemented behavior and the publication status of each package name.

Brand and Compatibility

PIO Agent is the product and Codex Plugin name. PlatformIO MCP is the underlying MCP runtime and the compatibility identity used by existing installations. The Codex plugin ID, marketplace ID, configuration keys, and skill namespace remain platformio-mcp. On npm, platformio-mcp is the canonical package while pio-mcp and pio-agent are thin compatibility packages that delegate to it. The canonical package also installs both platformio-mcp and pio-agent executable names. This lets existing consumers upgrade without migration while new users see PIO Agent throughout the interface.

It exposes PlatformIO workflows for board discovery, project setup, build, flash, monitor, diagnostics, and task orchestration through:

• an MCP server adapter
• a first-class CLI adapter (platformio-mcp / pio-agent)
• an optional local dashboard for visibility and control

MCP is one adapter. PlatformIO is the first backend.

Quick start (recommended)

npm install -g platformio-mcppio-agent devicespio-agent build --project-dir . --json

No server, no client configuration, nothing left running between commands.

Using it with a coding agent

Install the skills and point your agent at the pio-agent binary. The pio-manager skill documents every command.

MCP server (optional)

MCP remains fully supported for clients that prefer it:

claude mcp add platformio -- node /path/to/platformio-mcp/build/index.js

Note that each session that connects this way runs its own server process. The CLI above avoids that. See Manual MCP Config below for other hosts.

Agent-First Capabilities

• Project readiness validation (agent_validate_project)
• Rich build diagnostics with structured error taxonomy (agent_build_diagnose)
• Board-aware GPIO safety audits (agent_safe_pin_audit)
• Flash + monitor + runtime assertions (agent_flash_monitor_verify)
• Persistent workflow artifacts in .pio-mcp-workspace/ (lastAgentReport.json, boardReport.json)
• Board intelligence reports (agent_generate_board_report)
• Policy profile introspection (get_policy_status)
• Exact project/environment/device bindings (agent_resolve_target)
• Bounded, cursor-based serial health checks (agent_monitor_health)
• Idempotent task cancellation and compact history (cancel_task, list_task_history)
• Read-only approval status for agents; approval remains human-controlled

All risky operations still honor policy and approval rules.

Quick Start

1. Run the dashboard

npx platformio-mcp dashboard --serve

2. Use the CLI

All three npm names are supported. platformio-mcp is canonical; pio-mcp and pio-agent install the same CLI through thin dependency-only aliases:

npx platformio-mcp --helpnpx pio-mcp --helpnpx pio-agent --help

npx --package platformio-mcp pio-agent devicesnpx --package platformio-mcp pio-agent boards --filter esp32npx platformio-mcp init --board esp32dev --framework arduino --project-dir ./firmwarenpx platformio-mcp build --project-dir ./firmwarenpx platformio-mcp flash --project-dir ./firmware --port autonpx platformio-mcp monitor --project-dir ./firmware --port auto --expect BOOT_OK --timeout 30npx platformio-mcp agent-validate --project-dir ./firmwarenpx platformio-mcp agent-build-diagnose --project-dir ./firmwarenpx platformio-mcp agent-safe-pin-audit --project-dir ./firmware --board esp32devnpx platformio-mcp agent-flash-monitor-verify --project-dir ./firmware --expect-all BOOT_OK --reject-patterns "Guru Meditation,Brownout detector,WDT reset" --timeout 45npx platformio-mcp agent-last-report --project-dir ./firmwarenpx platformio-mcp policy-status --project-dir ./firmwarenpx platformio-mcp task-status <task-id>

Use --json for machine-readable output:

npx platformio-mcp build --project-dir ./firmware --json

3. Install into AI hosts

npx platformio-mcp install --clinenpx platformio-mcp install --claudenpx platformio-mcp install --vscodenpx platformio-mcp install --antigravitynpx platformio-mcp install --codex

4. Install the full Codex Plugin

The PIO Agent Codex Plugin adds the bundled MCP runtime, focused embedded skills, secure in-app dashboard flow, and monitoring-automation guidance. Codex requires one stable plugin identifier, so install selectors and skill namespaces remain platformio-mcp; the installed product is shown as PIO Agent. From a clone:

npm installnpm --prefix web installnpm run plugin:buildnode build/cli.js install --codex-plugin

From npm:

npx -y platformio-mcp install --codex-plugin

Start a new Codex task after installation. The legacy install --codex command remains available for MCP-only configuration. See the full Codex Plugin guide for update, uninstall, browser fallback, policy, automation, and rollback details.

The plugin release gates run on Windows, macOS, and Linux, exercise the authenticated dashboard in Chromium, validate the bundled runtime and public tool registry, and keep physical-board evidence in a separate manual workflow. That workflow uploads only bounded, sanitized evidence; raw hardware logs stay on the self-hosted runner. See the release and validation guide.

For headless verification and status inspection, the same CLI also provides plugin validate, target-resolve, monitor-status, monitor-health, task-history, approval-status, and pending-approvals. Run platformio-mcp --help for bounded options and JSON output support.

Manual MCP Config

{  "mcpServers": {    "platformio": {      "command": "npx",      "args": ["-y", "platformio-mcp", "serve"]    }  }}

On Windows, use npx.cmd if your host requires explicit shim resolution.

Core Capabilities

• Board and device discovery for PlatformIO-supported hardware
• Project initialization and config inspection
• Build, upload, monitor, and background task polling
• Structured diagnostics for build/upload/serial failures
• Safety and policy guardrails (approval gates, audit logs, redaction)
• Dashboard visibility for commands, logs, locks, and safety state

Safety Model

PIO Agent enforces policy decisions across CLI and MCP flows.

• Actions can be allow, deny, or requires_approval
• Risky operations (for example firmware upload/reset paths) require explicit approval
• All actions can be audited
• Secrets are redacted in exposed log streams

Policy profiles can be selected per-project via .pio-mcp-policy.json:

{  "profile": "flash_requires_approval"}

Supported profiles:

• read_only
• build_only
• monitor_only
• flash_requires_approval
• lab_runner (explicit, expiring unattended-lab policy required)
• lab_admin

CLI approval workflows:

npx platformio-mcp approvals --status pending --jsonnpx platformio-mcp approve <approval-id> --jsonnpx platformio-mcp deny <approval-id> --json

Codex Usage

Codex-facing docs and prompt cookbook:

• Codex Usage Guide
• Codex Prompt Cookbook

Documentation

Getting started:

• LLM Installation Guide

Guides and references:

• Agent Customization Guide
• MCP Server Command Reference
• Troubleshooting Guide
• Agent-First Embedded Workflow
• Competitive Positioning
• MCP-U Integration Template
• Agent Skills Directory

Specifications:

• PIO Agent Design Specification
• Web UX Design Specification
• Development Guide

Development

Prerequisites:

• Node.js >= 20 (including the bundled direct-serial runtime)
• PlatformIO Core CLI (install guide)

Local setup:

git clone https://github.com/jl-codes/platformio-mcp.gitcd platformio-mcpnpm installnpm run buildnpm run testnpm run smoke-test

CI/CD test tiers:

• npm run test:ci:unit runs unit/component coverage used in cross-platform CI.
• npm run test:e2e:ci runs CI-safe end-to-end tests for agent workflows and CLI wiring.
• .github/workflows/ci.yml runs typecheck, tests, and package smoke checks on pull requests/pushes.
• .github/workflows/hardware-e2e.yml is a manual self-hosted-runner workflow for one explicitly confirmed physical-board write, bundled-plugin protocol checks, post-flash identity/serial assertions, cleanup proof, and sanitized evidence.

Contributing

Contributions are welcome.

• Open an issue for bugs or feature requests
• Submit a pull request with tests when applicable

License

MIT. See LICENSE.

Firmware crash and size analysis

The MCP tools decode_backtrace and size_report analyze an explicitly selected project/environment using its registered GNU toolchain. Metadata and memory checks require build permission because they can execute project scripts. Reports identify the exact ELF, retain unresolved crash addresses, and distinguish PlatformIO memory usage from GNU estimates. They do not prove which firmware is on a device. See firmware analysis for inputs, approval behavior, evidence and current limits.