MCP server for Matter protocol device control and commissioning, providing tools to interact with Matter devices via Python Matter Server WebSocket interface and chip-tool commands
Matter MCP Server exhibits significant gaps in definition quality. While 15 tools are present with basic descriptions, most descriptions are under 100 characters and lack LLM-optimization guidance. Parameter descriptions are present but minimal (averaging 30-50 chars). Input schemas are visible in the source but lack detail: most parameters have type and description, but lack enums for constrained inputs, ranges for numeric values, and format specifications. Output schemas are not documented. Error handling is minimal, most tools return generic success/error responses without recovery guidance or error classification. No tool demonstrates idempotent design patterns or confirmation steps for destructive operations. The server lacks comprehensive documentation of what fields agents should expect in responses, forcing agents to reason about return structure. Example: commission_with_code returns a basic status dict but doesn't clarify the structure of the entire response or what downstream tools can use.
Commission a new device using QR code or manual pairing code.
Commission a new device using QR code or manual pairing code. Args: code: The QR Code or Manual Pairing Code (can be string or integer)
Send a command to a device.
Get a list of commands supported by a cluster Args: cluster_identifier: Either cluster name (e.g., 'DoorLock') or cluster ID (e.g., 257) Returns: List of command names supported by the cluster Raises: ValueError: If cluster cannot be found or XML cannot be parsed
Get the cluster Id from a cluster name Args: cluster_name: Name of the cluster (e.g., 'Identify', 'Descriptor') Returns: Cluster Id as an integer
Get the cluster name from a cluster Id Args: cluster_id: Cluster Id (e.g., 3 for Identify cluster) Returns: Cluster name as a string
Output schemas not documented. No tool describes what fields the response contains, their types, or what downstream tools can use. Example: commission_with_code returns [{"status", "message", "node_id"}] but the docstring doesn't specify this structure, forcing agents to infer.
Destructive operations (write_attribute, device_command) lack confirmation or dry-run patterns. No tool prevents accidental modification of critical device state.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 11 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 39 | - | v1 |
Get the device type Id from a device name Args: device_name: Name of the device type (e.g., 'Battery Storage', 'Cooktop') Returns: Device type Id as an integer
Get the device type name from a device type Id Args: device_id: Device type Id (e.g., 24 for Battery Storage) Returns: Device type name as a string
Get essential information about a specific node.
Get all commissioned nodes, returning only the descriptor (29) and basic information (40) cluster attributes.
Read an attribute from a node.
Search Matter protocol documentation on mattercoder.com Args: query: Search query related to Matter protocol (e.g., 'chip-tool', 'commissioning', 'clusters') Returns: List of dictionaries containing search results with titles and URLs
Set Thread credentials for device commissioning.
Set WiFi credentials for device commissioning.
Start listening for node events and changes.
Write an attribute value to a node.
Error handling is minimal. All tools return generic {"status": "error", "message": str(e)} responses. Errors don't classify as retryable, user-fixable, or fatal. No recovery guidance. Example: read_attribute might fail due to network timeout (retryable) or invalid node_id (user-fixable), but the agent can't tell the difference.
Parameter descriptions lack actionable constraints. Example: 'attribute_path' in read_attribute is described as 'The attribute path to read' but doesn't explain format (e.g., '0/29/0' vs endpoint/cluster/attribute). LLMs cannot infer valid input format.
No enum constraints on parameters with bounded choice spaces. Example: get_cluster_commands accepts 'cluster_identifier' as string|integer but doesn't specify valid cluster names or ID ranges. LLMs might pass arbitrary strings.
Parameter dependency undocumented. Example: device_command accepts endpoint_id, node_id, cluster_id, command_name, but the description doesn't explain dependencies (e.g., is endpoint_id required only for certain cluster_ids?) or provide examples of valid command structures.
Pagination not addressed for list-returning tools. get_nodes and get_node return potentially large attribute dictionaries but lack offset/limit parameters or documentation of maximum result size. No 'total count' or 'next_cursor' returned.
Description lengths are suboptimal for LLM selection. Most descriptions are 40-80 characters; baseline for A+ tools is 50-200. Example: start_listening description is only 42 chars ('Start listening for node events and changes.'), leaving ambiguity about what 'listening' entails (does it block? return a stream? set up a callback?).
Tool composition misses chaining context. Example: commission_with_code sets network credentials via set_wifi_credentials first, but the return value of commission_with_code doesn't include a 'requires_commissioning' flag or status that would tell an agent 'you must call set_wifi_credentials before this'.
Idempotency unclear for write operations. write_attribute and device_command don't document whether repeated calls with identical parameters are safe (idempotent) or risk side effects. Agents may retry on ambiguous errors and cause duplicate writes.