Reading a state is the easy half. From the second call in a session the response carries the previous action, the previous and current scores, and a plain boolean for whether stress actually fell — which is the piece an agent needs to change strategy rather than repeat itself. Video and audio stay on the device; the server takes derived signals, not recordings.
A hosted MCP server that takes sensor signals (heart rate, HRV, voice tone, expression, text sentiment) and returns a unified state with a machine-readable action the agent can follow. It is also reachable as an A2A agent for agent-to-agent use.
- `get_human_state` — a stress state, a suggested action such as maintain, simplify, de-escalate or pause, and the closed-loop feedback field
- `ingest` — send biometric signals and get the unified state back, with a subject id if you want trigger memory
- `get_trigger_memory` — which topics are still causing stress and which have resolved over time
- `get_session_history` — timestamped state history with a rising, falling or stable trend
- `request_api_key` and `check_api_key_status` — the agent asks for your real email, you click the verification link, it collects the key itself
- 6 tools in total; the two key-provisioning ones need no authentication
- The same capabilities as A2A skills for agent-to-agent calls
An MCP client that speaks Streamable HTTP — point it at https://mcp.nefesh.ai/mcp. No local install. Authentication is an API key in the X-Nefesh-Key header, or the agent can self-provision one: connect with no key, ask it to request a key with your real email address, and click the verification link. The free tier covers 1,000 calls a month with all signal types and no credit card. MIT licensed.
One command — npx -y mcp-remote https://mcp.nefesh.ai/mcp