On-device human sensing for AI: EDNA ID continuity and live human context via MCP.
Scored across 18 tools
The read-side tools overlap heavily: get_live_brief explicitly bundles proof + context + writing process + guidance, yet get_human_proof, get_human_context, get_writing_process and get_signal_vector are also exposed individually, forcing the agent to guess granularity. Similarly, edna_identity_verify vs edna.id.verify_resume and edna.submit_proof vs edna.id.feed have subtly different but poorly delineated roles. Descriptions help somewhat, but boundaries require careful reading rather than being self-evident.
Three conventions coexist: dot-namespaced `edna.get_*` namespace tools, flat snake_case `edna_identity_create` / `edna_identity_verify`, and `edna.id.*` chain tools. The verb placement also varies (get_human_context vs id.feed vs submit_proof). An agent cannot predict a tool name from its action without consulting the list.
Eighteen tools is on the heavy side, and the sensing half could be collapsed (get_live_brief already subsumes three of the individual getters). However, the identity half covers a genuinely broad cryptographic lifecycle, so the count is not gratuitous. Borderline but defensible.
The identity chain is unusually well covered (genesis, create, verify, resume gate, feed, memory, integrity, status, export, revoke), and the sensing side covers submission and all read channels. Gaps are minor: there is no tool to inspect or manage consent scope even though several tools require scope.id/scope.mcp, and no session lifecycle (start/stop) control.
Time-aware human context: dynamics, tempo trend, pauses, cautious indicators with limits.
Inputs: session.
Current human proof { confidence, classification } for this session.
Inputs: session.
One-call live brief for continuous perception: current proof + human context + writing process + response guidance in a single pull. Call before each answer while an EDNA session is streaming.
Inputs: session.
Adaptive Response Layer (production): overridable mode (FLOW/FOCUS/EXPLORE/EXECUTE/REFLECT) from consented dynamics. The humanness score is NEVER an input.
Inputs: session.
Quantized 16-channel behavioral signal vector of this session.
Inputs: session.
How to independently verify the engine (open protocol).
Inputs: session.
Writing-process channel (production): HOW the current message came to be — typing bursts, thinking pauses (the reason is never inferred), correction runs, tempo dynamics and a quantized 12-step timeline. Never content, never key identity.
Inputs: session.
EDNA ID (production): seal the device-signed identity anchor — only after edna.id.genesis and ≥3 minutes of full human verification on the Chladni hash chain. The seal is a device-key signature; it can only be created on the user's device (this endpoint validates it).
Inputs: session.
EDNA ID (production): verify a handle against the genesis handle commitment + live Chladni pattern vs account vault. Three consecutive 100% matches authenticate the session into the EDNA ID account (behavioral login); any miss resets the streak.
Inputs: handle (required), session, signals.
EDNA ID (production): local state export — no private key, no raw signals, no plaintext handle.
Inputs: session.
EDNA ID (production): feed latest proof into the chain; grows only on full human verification.
Inputs: session.
EDNA ID (production): device-bound EDNA ID (ECDSA-P256 keypair, signed handle, genesis from latest proof). Consent REQUIRED: scope.id === true. Continuity proof — never a real-world identity.
Inputs: handle (required), consent (required), session.
EDNA ID (production): verify signatures + linkage over full log and checkpoint chain.
Inputs: session, checkpointsOnly.
EDNA ID (production): memory layer live | session | longterm (vs the user's OWN anchor).
Inputs: layer, session.
EDNA ID (production): real LOCAL deletion — key, vault, model, log destroyed.
Inputs: session.
EDNA ID (production): chain status + measured identity-confidence curve (99%+ = research target).
Inputs: session.
EDNA ID (production): resume gate — reference-window match (>=0.85), never last-hash alone.
Inputs: session.
Submit the current human proof (score, classification, quantized signals, hash pattern, optional context/reading/process). Consent REQUIRED: scope.mcp === true. Never raw biometrics.
Inputs: consent (required), confidence (required), classification (required), context, +6 more.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Cross-device AI memory with encrypted activity capture and context handoff between AI tools
Fuses biometric signals into a stress score (0-100) for AI adaptation. MCP + A2A native.
EUDI Wallet MCP — EU Digital Identity Wallet under eIDAS 2.0 for AI agent authentication via
Private cross-media memory for AI assistants: recommendations, progress and controlled actions.
Add one secure layer between your agents and this server.
