Skip to main content
Glama

Server Configuration

Describes the environment variables required to run the server.

NameRequiredDescriptionDefault

No arguments

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

CapabilityDetails
tools
{
  "listChanged": false
}
prompts
{
  "listChanged": false
}
resources
{
  "subscribe": false,
  "listChanged": false
}

Tools

Functions exposed to the LLM to take actions

NameDescription
list_ordersA

List new (held-out) orders with their fields, so you can choose one to predict or explain.

Args: start: index of the first order to list. count: how many orders to list (keep this small).

predictA

Predict the target value for one order with the full model (all past orders in context).

Give either order_id (a held-out order, see list_orders) or fields (a dict of field values for a new order). Returns the predicted value, the model's confidence and the runners-up.

Args: order_id: index of a held-out order. fields: field values for a new order, as an alternative to order_id. model: "fast" (TabPFN-3.5-Fast) or "base" (TabPFN-3.5). Both run locally.

explainA

Find the past orders that the model's prediction seems to rest on (the precedents).

Takes the most similar past orders as candidates, repeatedly hides a random half and re-asks the model, then scores each candidate by how much it raises the model's confidence. Slow: roughly one minute on a GPU with the defaults. The result is UNVERIFIED; call verify with the returned explanation_id before presenting the precedents as the reason.

Args: order_id: index of a held-out order. fields: field values for a new order, as an alternative to order_id. model: "fast" or "base". candidates: how many similar past orders to consider. rounds: hide-and-ask rounds; about three per candidate gives stable scores. top_k: how many precedents to return.

verifyA

Check an explanation against the full model.

Removes the precedents from all past orders and asks the model again, then does the same with random non-precedent orders for comparison. Returns the fall in confidence, a verified true/false verdict, a strength label, and a plain-language summary to relay.

Args: explanation_id: id returned by explain. Preferred. order_id: alternatively, a held-out order; its latest explanation is used, or one is created. model: used only with order_id. "fast" or "base". n_random: number of random removals to compare against.

ledger_entriesA

Read the audit ledger: every predict, explain and verify call, in order.

Also reports whether the ledger's hash chain is intact (it breaks if an earlier entry was edited or deleted).

Args: last_n: how many of the most recent entries to return. order_id: only entries about this order. tool: only entries from this tool ("predict", "explain" or "verify").

Prompts

Interactive templates invoked by user choice

NameDescription

No prompts

Resources

Contextual data attached and managed by the client

NameDescription

No resources

TDQS

A3.9/5.0

Scored across 5 tools

Disambiguation5/5

Each tool has a clearly distinct role in the workflow: list held-out orders, predict a target, explain which precedents drive a prediction, verify an explanation, and read the audit ledger. Although predict and explain share order_id/fields arguments, their outputs and purposes are unambiguous, and the explain/verify relationship is explicitly described.

Naming Consistency3/5

The names mix conventions: list_orders is verb_noun, predict/explain/verify are bare verbs, and ledger_entries is a noun phrase. The set is still readable and understandable, but it does not follow a predictable single pattern.

Tool Count5/5

Five tools is well-scoped for a specialized prediction-with-precedents server. Each tool earns its place in the pipeline, and there is no redundant or filler tool.

Completeness4/5

The surface covers the core lifecycle: list examples, predict, explain, verify, and audit. A minor gap is the lack of a direct get_order tool or a way to retrieve arbitrary order details beyond listing, but agents can reasonably work around this with list_orders.

Maintenance

ActivitySlowing
ResponsivenessNo recent community issues