PIO Agent README
The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the PIO Agent listing page.
PIO Agent
PIO Agent is the open-source, agent-first hardware execution layer for embedded development, built on the PlatformIO MCP runtime.
3.1.0 development status
The parity branch prepares an unreleased 3.1.0 version with optional reference tool aliases, owned serial workflows, retained flash, OTA and expanded distribution candidates. Existing published installations remain 3.0.0. See the unreleased changelog, compatibility guide, and distribution readiness for implemented behavior, remaining acceptance, and which package names are prepared versus published.
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.
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
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:
Use --json for machine-readable output:
3. Install into AI hosts
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:
From npm:
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 44-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
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, orrequires_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:
Supported profiles:
read_onlybuild_onlymonitor_onlyflash_requires_approvallab_runner(explicit, expiring unattended-lab policy required)lab_admin
CLI approval workflows:
Codex Usage
Codex-facing docs and prompt cookbook:
Documentation
Getting started:
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:
Development
Prerequisites:
- Node.js >= 20 (including the bundled direct-serial runtime)
- PlatformIO Core CLI (install guide)
Local setup:
CI/CD test tiers:
npm run test:ci:unitruns unit/component coverage used in cross-platform CI.npm run test:e2e:ciruns CI-safe end-to-end tests for agent workflows and CLI wiring..github/workflows/ci.ymlruns typecheck, tests, and package smoke checks on pull requests/pushes..github/workflows/hardware-e2e.ymlis 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.