Odin Assistant

Voice and chat. With a receipt for every prompt.

The assistant inside your workspace. XTTS + Whisper for voice; Anthropic, OpenAI, OpenRouter, or local MLX for chat. Same governance contract across modalities.

Request enterprise access See the workspace

BuilderOperator

Modalities

Voice and chat. One contract.

The assistant accepts speech or text and replies in kind. Voice is XTTS for output and Whisper for input, GPU-accelerated. Chat routes through whichever provider you've configured for the tenant.

modality · voice

GPU

→Browser-native — microphone capture in the workspace tab; no native app.
→Low-latency — GPU inference keeps end-to-end under 800ms.
→5 environments — voice surfaces across the suite, every audit trail.

TTSXTTS v2

STTWhisper

modality · chat

multi-provider

→Anthropic / OpenAI / OpenRouter / local MLX — per-tenant provider keys.
→Cost-routed — rule 27 tier ladder picks cheapest capable for each task.
→Fallback chain — provider failure routes to next tier, audited.

providers4+

tiersT0 → T6

Routing

Cheapest capable. Audited per call.

A 7-tier cost ladder picks the right model for each task — T0 local Ollama, T2 Haiku, T3 Sonnet, T6 frontier. Every routing decision lands in your audit trail with the rationale, estimated cost, and fallback chain.

Tier ladder

T0 — Ollama local$0
T0.5 — MLX local$0
T1 — Factory plan$20/mo
T2 — Haiku 4.5$1 / 1M
T3 — Sonnet via OR$3 / 1M
T5 — Codex CLImetered
T6 — Opus / GPT-5$5 / 1M

routing-decisions.log · last 5

audited

summarization · digestT2$0.003
multi-file refactorT5$0.18
architecture decisionT6$1.42
classificationT0$0
unit test genT5$0.04

Audited

Every prompt. Every reply. Captured.

The audit trail doesn't sample. Every prompt the assistant receives + every response it returns lands in BrainDB with token usage, model, provider, and per-call cost. Forensic by construction.

brain/audit/assistant.log

decisions

10:14:02user:operator→voice.prompt#a8f2
“How do I add OAuth2 to the customer portal?”
10:14:03agent:assistant→router.decide#b14c
“Routed to T3 Sonnet — multi-file context required”
10:14:07agent:assistant→chat.reply#c903
“1,247 input tokens · 482 output · $0.0083”
10:14:11agent:assistant→tts.synthesize#d6e1
“XTTS v2 · 612ms · GPU”
10:15:30user:operator→voice.prompt#e7a8
“Use authlib not passport”
10:15:31agent:brain→memory.write#f211
“Decision recorded: chose authlib over passport (rationale 18 chars)”