mcp-laya
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| LAYA_MODELS | No | Allowlist of model checkpoints that may be loaded. | |
| LAYA_REDACT_STATE | No | If true, input text is redacted from the JSON audit log. | true |
| LAYA_DEFAULT_MODEL | No | Default model to use, e.g. 'auto'. | auto |
| LAYA_MAX_QUESTIONS | No | Maximum number of questions per request. | |
| LAYA_ALLOW_DOWNLOAD | No | Allow downloading model checkpoints on first run. Off by default. | false |
| LAYA_MIN_CONFIDENCE | No | Minimum confidence threshold; lower-confidence answers are flagged. | |
| LAYA_MAX_INPUT_CHARS | No | Maximum number of input characters per request. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| decideA | Answer several typed questions about a state in one local forward pass. Use this to offload System-1 judgment calls from the main model: classification, routing, scoring and yes/no gates over text, an email, a ticket or a JSON object — fast, local, and without hallucination. Each question is one of three types:
Returns every answer with its calibrated |
| classifyB | Assign the single best category to a state (one Provide either |
| scoreB | Rate a state on an ordinal scale (one
|
| checkA | Answer a single yes/no/unknown question about a state (a Returns |
| triageA | Run a ready-made set of decisions over a state using a laya preset. presets: triage (department/urgency/sentiment…), email, moderation (safety categories), guard (prompt-injection / policy gates). Returns all answers with confidence. |
| detect_languageB | Detect the script and language of a text (sub-millisecond, no model forward pass). |
| explain_routingA | Explain which checkpoint the Router would use for this state, without running inference. |
| list_modelsA | List the laya checkpoints this server may use, the default, device and offline status. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 8 tools
classify, score, and check are single-question wrappers around decide's choice/score/noul modes, so an agent choosing between decide and a specialized tool may face ambiguity for a one-off task. Their descriptions clearly separate the single-vs-multiple use cases, but the underlying semantics overlap.
All tool names are lowercase and start with an imperative verb, with compound names following verb_noun (detect_language, explain_routing, list_models). The single-word verbs (decide, classify, score, check, triage) are a minor stylistic deviation from the compound pattern but remain predictable.
Eight tools is a reasonable size for a decision-model server and stays inside the ideal 3-15 range. The count is slightly padded, though, because classify/score/check are specialized versions of decide's three question types.
The set covers general and specialized decisions, ready-made triage presets, language detection, routing explanation, and model listing—no obvious dead ends for inference workflows. Minor gaps include no queryable preset list or way to register custom presets, but these can be worked around.