agent-runtime · test client

disconnected

llm: idle

tenant (none picked) group —

session · dev only ▾

tenant

loading…

Sent as ?tenant_id=… on WS upgrade. Picks the tenant's agents, tools, and backend URL.

bearer JWT

Sent as ?token=… on WS upgrade. The runtime relays it to tools with auth_ref: "session".

org id

Sent as ?org_id=…. Seeded into the runtime variable bag, so tool params with source: "runtime_variable" and variable_ref: "org_id" resolve here. Future: derive from verified JWT claims.

routing cache · L2 —

click refresh to load

auto-refresh 5s

agent config · LLM tuning defaults ▾

Sent in llmConfig on session.start. For voice mode these values are the LOWEST-precedence source — Mongo tenant_llm_configs > per-language default > wire. For text mode they apply directly (REST stateless).

temperature 0.2

max output tokens

top_p unset

top_k unset

top_p and top_k are sent only when non-zero. Leave at 0 to let the model pick its defaults.

audio pipeline · voice only ▾

RNNoise (fast WASM denoise)

RNNoise intensity 1.0

DeepFilterNet (neural denoise — needs ONNX model)

DeepFilterNet intensity 1.0

Record per-turn audio in chat (in-memory only)

Toggles take effect on the NEXT call. RNNoise needs no model file (bundled in WASM); DeepFilterNet gracefully no-ops if audio-models/deepfilternet3.onnx is missing.
Intensity is a wet/dry mix: 1.0 = full denoising (default), 0.5 = half-strength, 0.0 = bypass. Lower it if the denoiser sounds "underwater" or chews up sibilants.

chain state · live no session

status— agent—

(no session yet)

intent queue · paused + waiting empty ▾

last reason—

(no pending intents)

initial context · user pre-fetch ▾

preset

intent

No user pre-fetch — the greeter will collect name + DOB during the call.

mode model text model voice model voice language