Blea MCP Server

BLEA

BLEA — Bluetooth Low Energy Automation — is a universal Agent Plugin for safe, deterministic BLE work. It gives AI agents a portable Skill, a machine-readable CLI, a local MCP server, guarded automation, and offline evidence capture, diff, and replay.

The project is designed for the BLE work agents repeatedly need: checking OS Bluetooth access, finding the right nearby device, discovering GATT services, preserving and comparing raw byte evidence, reading and subscribing before changing state, and making writes only after explicit authorization.

The ble executable is the shared runtime. Regular subcommands use the CLI surface; ble mcp starts the stdio MCP server. The MCP protocol is therefore a first-class interface, not a second tool hidden inside an unrelated CLI product.

Install

BLEA requires Python 3.10 or newer. Bluetooth access is provided by the operating system through Bleak.

The version-aligned 0.6.4 commands below are available once the matching GitHub and PyPI release is published:

uv tool install "blea==0.6.4"
ble --help

For Codex, install the public Git marketplace package from the matching immutable release tag after the runtime:

codex plugin marketplace add Nitmi/blea --ref v0.6.4
codex plugin add blea@blea

To install only the portable Agent Skill with the Skills CLI:

npx skills add Nitmi/blea --skill ble

The Skill teaches the workflow and safety policy; it does not install the Python runtime. Install blea from PyPI as shown above before asking an agent to access live Bluetooth hardware or start the local MCP server. The indexed Skill page is skills.sh/nitmi/blea/ble.

Start a new Agent task after installation. The marketplace package supplies the BLE Skill and MCP declaration; it does not install the Python runtime or operating-system Bluetooth permissions. See docs/installation.md for the release-tag, legacy 0.6.0 bridge, and update procedure.

From a checkout, install the command runtime directly from the working tree:

uv tool install --editable .
ble --version

For repository development:

uv sync --extra dev
uv run ble --help

The Python runtime and Agent Plugin are two related installation units: PyPI installs the ble executable, while an Agent Plugins 1.0.0 client loads this repository directory to discover plugin.json, mcp.json, and skills/ble. Install the runtime first so the plugin's local stdio MCP server can find ble on PATH. See docs/installation.md for source, plugin, update, permissions, and verification paths.

CLI

ble doctor --json
ble scan --timeout 8 --json
ble inspect --device "id:AA:BB:CC:DD:EE:FF" --json
ble probe --device "id:AA:BB:CC:DD:EE:FF" --max-reads 16 --json
ble probe --device "id:AA:BB:CC:DD:EE:FF" --max-reads 16 --read-offset 16 --json
ble read --device "id:AA:BB:CC:DD:EE:FF" --characteristic 2a19 --json
ble subscribe --device "id:AA:BB:CC:DD:EE:FF" --characteristic 2a37 --duration 15 --jsonl
ble observe --device "id:AA:BB:CC:DD:EE:FF" --duration 10 --jsonl
ble diff before.blea.jsonl after.blea.jsonl --json
ble replay capture.blea.jsonl inspect --json
ble replay capture.blea.jsonl read --characteristic 2a19 --json
ble exchange \
  --device "id:AA:BB:CC:DD:EE:FF" \
  --write-characteristic 12345678-1234-1234-1234-1234567890ab \
  --notify-characteristic 87654321-4321-4321-4321-ba0987654321 \
  --text ping --duration 5 --allow-write \
  --confirm-device "AA:BB:CC:DD:EE:FF" --jsonl

Device names are accepted only when exactly one observed device has that name. Prefer the platform identifier returned by scan; on macOS this is normally a UUID rather than a MAC address.

Every binary value includes length, hex, base64, and replacement-safe UTF-8 representations. JSON errors use stable reasons and exit codes suitable for agent recovery.

probe is paginated. Follow next_read_offset until it is null. Page counts and characteristic-level failures live under read_page; ok=true means the page executed, while read_page.has_failures tells you whether any characteristic failed. Every result includes a small profile_summary. The CLI includes the full GATT profile by default and supports --no-include-profile; the MCP tool omits it by default to keep repeated pages compact.

Connection commands report timeout_scope=per_backend_operation. Their timeout applies separately to discovery, connection, and each GATT operation, rather than bounding the total command time. GATT entries include uuid_namespace. BLEA reports library descriptions only for canonical Bluetooth Base UUIDs, avoiding false standard names inferred from the leading bytes of custom 128-bit UUIDs.

observe discovers notify/indicate characteristics and watches them over one connection for a bounded duration. Omit --characteristic to select all event-capable traits, or repeat it for explicit selection. The result separates subscription failures, cleanup failures, and notification events; a quiet window is evidence only for that sample window.

Save a complete read-only evidence package after diagnosis:

ble capture \
  --device "id:AA:BB:CC:DD:EE:FF" \
  --output capture.blea.jsonl \
  --max-reads 128 \
  --observe-duration 10 \
  --redact-identifiers \
  --json

capture records the advertisement, normalized GATT profile, each readable-characteristic result, bounded notifications, operation errors, and a validated final summary. It performs no writes, pairing, or configuration changes and replaces the destination only after an atomic close. The JSONL file is the authoritative artifact; the command's JSON result is a compact summary. The Evidence Format v1 contract and deterministic validator live in docs/evidence-format-v1.md.

Compare two complete evidence packages without a Bluetooth adapter or device:

ble diff before.blea.jsonl after.blea.jsonl --json
ble diff before.blea.jsonl after.blea.jsonl --fail-on-change --json

diff validates both Evidence Format v1 inputs, projects them into stable BLE semantics, and emits sorted JSON Pointer changes. It ignores capture IDs, timestamps, runtime metadata, sampling duration, and RSSI movement within 5 dBm by default. Use --strict-rssi for exact RSSI comparison. Device identifiers must match unless an intentional cross-device comparison uses --allow-different-devices. Binary payloads are atomic changes, so one changed value does not produce separate Hex, Base64, and UTF-8 noise. A normal difference exits successfully; --fail-on-change returns code 3 after printing the complete result for CI. The contract lives in docs/diff-format-v1.md.

Replay a complete evidence package without a Bluetooth adapter or physical device:

ble replay capture.blea.jsonl inspect --json
ble replay capture.blea.jsonl probe --max-reads 32 --json
ble replay capture.blea.jsonl read --characteristic 2a19 --json
ble replay capture.blea.jsonl observe --duration 10 --jsonl
ble replay capture.blea.jsonl run examples/replay-read-only.yaml --json

Replay reconstructs advertisements, GATT, reads, captured failures, subscription outcomes, and the notification timeline through the same read-only backend interfaces used for live devices. The default --speed 0 mode is immediate and deterministic; place a positive --speed before the operation to preserve recorded notification gaps at that multiplier. Missing observations return the stable replay_miss reason instead of an invented value or success. Replay never accesses a real adapter and never sends or simulates writes or exchanges. The full contract lives in docs/replay-format-v1.md.

Guarded writes

A write requires both --allow-write and an exact confirmation of the resolved identifier:

ble write \
  --device "id:AA:BB:CC:DD:EE:FF" \
  --characteristic 12345678-1234-1234-1234-1234567890ab \
  --hex 01 \
  --allow-write \
  --confirm-device "AA:BB:CC:DD:EE:FF" \
  --read-back \
  --json

YAML writes and exchanges add two more guards: the workflow must enable writes and each state-changing step must declare dangerous: true plus successful prerequisite steps. The policy must also carry an exact confirm_device identifier. See examples/guarded-write.yaml and examples/esp32-burst-exchange.yaml.

For protocols where a write triggers notifications, use exchange. It enables the notification subscription before performing the guarded write, then collects events for a bounded duration. The write and notify characteristics may be the same or different. This avoids the race created by launching standalone session subscribe and write operations concurrently.

For repeatable checks, an exchange step can assert the exact notification count, UTF-8 or Hex content, the final notification, and subscription cleanup. Run a guarded YAML workflow with the independent invocation gate:

ble run examples/esp32-burst-exchange.yaml --allow-write --json

Replace the example's device identifier and confirmation with the exact value returned by a fresh scan before running it.

MCP and Agent Plugin

Start the local stdio server directly:

ble mcp

The MCP surface includes one-shot tools, the offline ble_diff comparator, the offline ble_replay runner, and stateful session tools. Sessions let an agent connect once, inspect, read, observe, subscribe, perform a guarded request/notification exchange or write, and then disconnect. ble_exchange and ble_session_exchange atomically subscribe before writing. ble_session_list exposes active leases and ble_session_close_all provides explicit recovery. The server disconnects all sessions when the MCP client exits and reaps an inactive session after 120 seconds by default. Set BLEA_SESSION_IDLE_SECONDS to another positive duration, or 0 to disable idle reaping while retaining shutdown cleanup.

Close a known session with ble_session_close. Reserve ble_session_close_all for recovering an unknown session ID, a failed explicit close, or confirmed leaked state.

To expose the normal MCP BLE tools against one capture instead of hardware, launch a dedicated replay server:

ble replay capture.blea.jsonl mcp

Every replay-backed BLE operation result identifies the evidence and carries replay.read_only=true. Even if a client invokes an existing write or exchange tool with live-device authorization fields, the ReplayBackend rejects it.

This repository is itself an Agent Plugins 1.0.0 package:

blea/
├── plugin.json                        portable plugin identity
├── mcp.json                           portable local stdio MCP declaration
├── skills/ble/SKILL.md                portable Agent Skill
├── .codex-plugin/plugin.json          root ChatGPT/Codex package metadata
├── .mcp.json                          root ChatGPT/Codex MCP declaration
├── .agents/plugins/marketplace.json   public Git marketplace catalog
└── plugins/blea/                      installable Codex distribution mirror

The plugin configuration expects the ble executable to be installed on PATH. The portable specification distributes Skill and MCP metadata; native Bluetooth runtime installation and OS permissions remain platform responsibilities.

The repository also carries server.json, the official MCP Registry package metadata for the PyPI distribution. It declares a local stdio server and the mcp package argument; it does not advertise a hosted or remote BLE service. Releases publish this metadata to the official MCP Registry with short-lived GitHub OIDC credentials after the matching PyPI version succeeds. The namespace preserves the GitHub login's canonical casing because Registry authorization and package ownership markers are case-sensitive.

The root Dockerfile is a registry sandbox for MCP protocol introspection and adapter-free tools. It builds the version-aligned release checkout and starts ble mcp, which lets release-candidate CI run before that version exists on PyPI. It does not claim that a container can access the host's Bluetooth adapter; use the native host installation above for live BLE work.

The OpenAI Plugins Directory submission uses the portable Skill only. BLEA's MCP server stays local because live BLE requires the user's native host and adapter; it is not represented as a public remote MCP endpoint. Submission copy, limitations, starter prompts, and review cases are recorded in docs/openai-plugin-submission.md.

The public Codex distribution is registered at the immutable Git ref shown in the install section and installed with codex plugin add blea@blea. A pinned ref does not advance to another release when a marketplace snapshot is refreshed. To change releases, remove the configured marketplace, add it again at the new v<version> tag, reinstall the Plugin, and start a new Agent task. Do not treat installing the Python package alone as installing the Agent Skill or MCP declaration.

Platform status

Windows 11 is hardware-verified with the ESP32-S3 BLEA test server, including discovery, GATT, notifications, guarded exchange, capture, and replay. macOS and Linux currently pass the unit suite and replay the same real-device evidence in CI, but their native Bluetooth adapter paths have not been tested on physical hosts. Treat those two platforms as CI + replay verified; native BLE hardware unverified until a published acceptance report says otherwise.

See docs/platform-acceptance.md for the exact support tiers, privacy requirements, hardware acceptance procedure, and current evidence matrix.

Development

uv sync --extra dev
uv run python scripts/sync_codex_plugin.py --check
uv run python scripts/check_agent_package.py
uv run ruff check .
uv run ruff format --check .
uv run pytest
uv build --clear --no-create-gitignore
uv run python scripts/check_distribution.py dist

Unit tests use a fake BLE backend and do not require nearby hardware. CI runs on Windows, macOS, and Linux. Each CI job also runs examples/replay-read-only.yaml against the checked-in complete evidence fixture, providing an end-to-end CLI and Workflow smoke test with no adapter.

The root .codex-plugin, .mcp.json, and skills paths are the Codex Plugin source of truth. After changing them, run uv run python scripts/sync_codex_plugin.py; CI uses --check to reject drift. The repository-local Agent package checker validates both Codex Plugin roots, their MCP declaration, Skill metadata, and agents/openai.yaml without relying on a maintainer's personal Skill directory. The distribution checker also rejects missing marketplace files in the sdist and Plugin metadata in the Python wheel.

Real adapter support is tracked separately from CI; CI success is not a native hardware support claim.

Release history lives in CHANGELOG.md. Maintainers should follow docs/releasing.md for artifact, Trusted Publishing, Plugin, and post-publication gates.