TypeSafe MCP
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| TYPESAFE_MODEL | No | Model alias; legacy name supported | jev-latest |
| TYPESAFE_API_KEY | Yes | Required bearer credential | |
| TYPESAFE_BASE_URL | No | API base URL | https://api.typesafe.ai |
| TYPESAFE_MAX_RETRIES | No | Retries after the initial request | 2 |
| TYPESAFE_DEFAULT_MODEL | No | Official SDK-compatible model name | jev-latest |
| TYPESAFE_MAX_STATE_CHARS | No | Serialized state limit | 120000 |
| TYPESAFE_TIMEOUT_SECONDS | No | Per HTTP attempt timeout | 10 |
| TYPESAFE_MAX_REQUEST_BYTES | No | Whole request limit | 512000 |
| TYPESAFE_MAX_QUESTION_CHARS | No | Serialized question limit | 60000 |
| TYPESAFE_MAX_RESPONSE_BYTES | No | Provider response limit | 4194304 |
| TYPESAFE_RETRY_BACKOFF_SECONDS | No | Initial exponential backoff | 0.5 |
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
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| evaluateA | USE WHEN you need raw typed Jev answers for one or more bounded noul, choice, or score questions. DO NOT USE WHEN you need prose, code generation, arithmetic, dates, or a specialized primitive such as route or gate. |
| classifyA | USE WHEN you need exactly one label from a closed, unordered set and its probability distribution. DO NOT USE WHEN labels are ordered levels (use score), options are next actions (use route), or inputs are claims (use verify). |
| scoreA | USE WHEN you need a weighted rating over two or more ordered rubric levels. DO NOT USE WHEN choices are unordered categories (use classify), next actions (use route), or thresholded checks (use gate). |
| checkA | USE WHEN you need one bounded yes/no proposition and its probability. DO NOT USE WHEN you need to verify multiple named claims (use verify), combine checks with thresholds (use gate), or assess an entire object or review package (use review). |
| verifyA | USE WHEN you need to check multiple named claims against supplied evidence in one request. DO NOT USE WHEN you need one proposition (use check), an overall quality/risk review (use review), or a pass/review/fail threshold (use gate). Results are signals, not proof. |
| gateA | USE WHEN you need to combine multiple bounded checks and thresholds into pass/review/fail. DO NOT USE WHEN you need claim-by-claim evidence checking (use verify) or whole-object quality/risk assessment (use review). This is not authorization or a security boundary. |
| routeA | USE WHEN you need exactly one next action from a closed set of actions. DO NOT USE WHEN you need a topic/category label (use classify), an ordered rating (use score), or action execution. This suggests an action; it never executes it. |
| reviewA | USE WHEN you need an overall quality or risk review of a whole-object artifact such as a diff, plan, release, or test report against multiple checks. DO NOT USE WHEN you need claim-by-claim verification (use verify), a single proposition (use check), or next-action selection (use route). It returns signals and never edits files. |
| healthA | USE WHEN you need local Arbitype configuration or an explicitly requested live provider check. DO NOT USE WHEN you need a business decision; health does not evaluate state and does not make a network request unless live=true is explicit. |
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 9 tools
Each tool has a clearly distinct purpose, reinforced by explicit DO NOT USE cross-references that map boundary cases to the correct tool. Even the broadest tool, evaluate, is positioned as the raw-answer fallback rather than a competitor to the specialized distribution, threshold, or routing tools.
The tools all follow a consistent lowercase single-word imperative style, which is easy to predict and read. The only minor deviation is health, which is a noun rather than a verb, so the pattern is not perfectly uniform.
Nine tools is a well-scoped count for a typed evaluation and decision primitive library. Each tool represents a distinct decision mode, and none feel redundant, superficial, or missing.
The surface covers raw typed evaluation, classification, scoring, whole-object review, single and multi-claim verification, threshold gating, routing, and health checks. For the apparent purpose of safe, typed decision primitives, there are no significant gaps or dead ends.