nb-speak GitHub
nb-speak/1 · Universal IM Bubble & Artifact Protocol

How AI agents speak human.

Replace 300-word Markdown blobs with paced send_message beats, slide-out send_artifact cards, and ElevenLabs voice notes.

Raw LLM Blob (assistant.content) bubble_lint: 4 issues
markdown_structure · over_en_budget 180-word essay wall
nb-speak (send_message + send_artifact) bubble_lint: 0 warnings
3 bubbles · 1.5s cadence
DSH Live · :19081 vs :19082 Session 5 · Casual Friday Convenience Store Chat (:19081 vs :19082)

Live Bridge & Protocol Workspace

Connect to npx nb-speak serve (127.0.0.1:18090), bring an OpenAI/DeepSeek key in browser memory, or normalize raw Markdown walls offline.

Checking local bridge...
bubble_lint: 0 warnings

    Drop Into Any Stack in 30 Seconds

    Zero runtime dependencies. Works as an Agent Skill, an OpenAI/SSE proxy bridge, or a native DeepSeek Harness plugin.

    1. Install chat-human Agent Skill
    npx skills add NB-Corp/nb-speak
    2. Start OpenAI / SSE Bridge
    npx --yes github:NB-Corp/nb-speak serve --port 18090
    3. Mount in DeepSeek Harness (DSH)
    npx --yes github:NB-Corp/nb-speak-dsh --apply --profile web
    Capability nb-speak Raw Chat UIs (OpenWebUI / LobeChat) Prose Humanizers Plain TTS Readout
    Output unit 1–4 paced bubbles + send_artifact cards Single 300-word assistant.content blob Rewritten prose paragraphs Audio clone of full Markdown blob
    Code, SQL & tables Offloaded cleanly via send_artifact Crammed inline inside narrow bubbles Left untouched in text stream Reads backticks & URLs aloud
    Voice note physics Final-bubble complementary aside (<50% overlap) Full-message screen reader TTS None 100% verbatim text duplication
    Subagent tool safety Transparent send_message <-> agent_send_message remap Collides with agent orchestration tools N/A N/A
    Deterministic gate TS + Python bubble_lint (97.2%–100% pass rate) None Subjective prompt heuristics None
    Live DeepSeek Harness Capture (:19081 vs :19082)
    DeepSeek Harness side-by-side comparison