Vector Decisions MCP Server
# Vector Decisions MCP Server
MCP-native governance and action-assurance primitives for autonomous AI.
## Tools
- `gatri_trust_score` — legacy composite trust scoring.
- `assure_action` — evidence-backed, context-aware assurance for one agent action.
- `decide_action` — converts GATRI assurance + policy + permissions + context into `ALLOW`, `RESTRICT`, `REQUIRE_HUMAN`, or `DENY`.
- `risk_assessment` — deployment risk assessment.
- `compliance_check` — EU AI Act-oriented policy check.
- `kill_switch` — emergency stop directive.
## Core model
```text
agent × action × context × evidence
↓
GATRI
↓
Vector Decisions
↓
ALLOW / RESTRICT /
REQUIRE_HUMAN / DENY
```
The important primitive is **action assurance**, not a generic claim that an agent is safe. The same agent can be acceptable for one action and unacceptable for another.
## Local development
```bash
npm install
npm run build
npm start
```
The server uses stdio transport and is intended to be consumed by MCP-compatible clients.
## Status
Version 1.1.0. This is a protocol-oriented V0: deterministic local evaluation first; evidence ingestion, persistent assurance records, policy registries, signed attestations and remote APIs are subsequent layers.
TDQS
Scored across 6 tools
Each tool targets a distinct concern: trust scoring, emergency halt, compliance, risk assessment, action assurance, and decision-making. While trust and risk are related, they are conceptually different (governance quality vs. deployment risk), and the assurance/decision pair forms a clear pipeline without overlap.
All tool names follow a consistent snake_case pattern with clear verb/noun structure (e.g., kill_switch, compliance_check, decide_action). The use of the 'gatri' prefix in gatri_trust_score is a minor deviation but doesn't break the consistency.
Six tools is well within the ideal range and maps cleanly to the server's governance/decision-making domain. Each tool adds unique capability without redundancy, and the count feels appropriately scoped.
The tool surface covers the core lifecycle: assessment (trust, risk, compliance), action-level assurance, decision output, and emergency control. Minor gaps exist, such as policy management or audit trail retrieval, but these are not essential to the primary workflow and can be worked around.