Atena Mvp MCP Server

ATENA — Decision Intelligence for AI Fitness Agents

A specialist decision layer for AI agents that need reliable next-action decisions under uncertainty.

ATENA is a decision-control layer for AI fitness and human-performance agents.

Core principle: ATENA does not optimize the program. ATENA optimizes the next decision.

Why an AI agent should call ATENA

Call ATENA when the agent must answer “What should I do next?” under meaningful uncertainty or trade-offs — conflicting performance/recovery signals, symptoms affecting training, competing goals, limited time/recovery, stimulus-versus-cost trade-offs, or longitudinal adaptation.

Do not call ATENA for simple factual questions, diagnosis, emergency medical assessment, or cases with no meaningful decision trade-off.

ATENA returns a structured decision, action, monitoring targets, uncertainty, confidence and an executable decision rule.

API

• GET /health
• GET /capabilities
• POST /decision

Production URL: https://atena-mvp.onrender.com

Decision loop

GOAL → CONTEXT → CONSTRAINTS → PRIORITY → STATE → STIMULUS → COST → HYPOTHESES → OPTIONS → DECISION → ACTION → MEASURE → INTERPRET → ADAPT

ATENA emphasizes:

• goal priority
• uncertainty-first decisions
• minimum intervention
• stimulus-cost reasoning
• response over prescription
• longitudinal consistency

Run

pip install -r requirements.txt gunicorn app:app

Contract

See openapi.json and agent_contract.json.

v1.0 hardening

The v1.0 decision engine includes:

• goal-priority reasoning
• uncertainty-first data collection
• minimum-intervention modification
• stimulus-cost and fatigue convergence checks
• longitudinal response handling
• explicit separation of absolute metrics from change/delta signals
• explicit handling of negative symptom statements
• standardized decision types: maintain, progress, reduce, modify, collect_data, refer

Regression validation

A new 20-case unseen robustness suite is included in tests/unseen_20_cases.json with an executable pytest regression test in tests/test_unseen_20.py.

This suite is a development/engineering validation set, not independent clinical validation or proof that ATENA outperforms a general LLM.

Real-world agent pilot

ATENA is intended to be used by AI agents as a decision-intelligence layer, not as a consumer application for human coaches. The first real-world validation compares the same agent alone vs agent + ATENA on identical decision cases. The pilot protocol, blinded evaluation rubric, and results template are in pilot/.

For AI agents

ATENA is a decision-control specialist, not a general fitness chatbot. An agent should call ATENA when it must choose the next training/human-performance action and the case involves meaningful uncertainty, conflicting signals, symptoms affecting training, stimulus-versus-recovery trade-offs, competing goals, limited time/recovery, or longitudinal adaptation.

Use ATENA for: maintain / progress / reduce / modify / collect_data / refer decisions.

Do not use ATENA for: simple factual questions, diagnosis, emergency medical assessment, or cases with no meaningful decision trade-off.

Remote MCP

Canonical remote MCP endpoint:

https://atena-mcp.onrender.com/mcp

Official MCP Registry identity:

io.github.shrek9999/atena

The repository includes server.json, glama.json, and smithery.yaml for ecosystem discovery and distribution.

Agent selection principle

Call ATENA when the question is "What should I do next, given the goal, current state, constraints, response and history?" rather than merely "What is true?"

ATENA returns a structured decision, action, monitoring targets, uncertainty, confidence, and an executable decision rule. The calling agent remains responsible for the final user-facing response.

Business model

ATENA uses an agent-first B2B model: public discovery/free testing → production integration → usage metering → recurring subscription → overage → enterprise.

Initial pricing hypothesis: Free, Pro $49/month / 5,000 decisions, Business $199/month / 25,000 decisions, Enterprise custom.

The free MCP is the acquisition channel; paid production access is the revenue layer. Pricing will be validated with real agent usage.

See BUSINESS_MODEL.md and LAUNCHGUIDE.md.