Haiven API Documentation

AI workflow automation and orchestration

Web file manager for uploading AI models and browsing storage

MCP protocol server with embedded Whisper STT, 23 tools, and OpenAI-compatible audio API

MCP protocol server wrapping LiteLLM proxy for local LLM calls (GLM-4.7-Flash, Seed-36B) without Claude API quota

AI-optimized web scraping with JavaScript rendering, LLM-friendly markdown, and RAG pipeline integration

On-demand GGUF model gateway with OpenAI-compatible chat/completions APIs.

vLLM always-on GLM-4.7-Flash AWQ service with 200K context for fast mechanical and code-adjacent tasks.

vLLM always-on Bravo service serving qwen3.6-35b-a3b with 262K context for general-purpose completion.

vLLM always-on Huihui-Qwen3.6-27B-abliterated-BF16 service (runtime FP8, 200K context) for creative writing and low-refusal prose.

vLLM always-on Qwen3-Embedding-4B BF16 service with 8K context for semantic embeddings. Embeddings-only — use /v1/embeddings, not /v1/chat/completions.

vLLM always-on Qwen3.5-27B FP8 service on Delta GPU with 128K context, tool calling, and structured output. Thinking disabled — optimized for JSON generation and classification.

vLLM always-on MedGemma 27B Text IT FP8 service with 32K context for medical text comprehension, clinical reasoning, and biomedical QA. Text-only (no vision). Shared Delta GPU.

vLLM on-demand MiniMax-M2.5 AWQ Q4 229B MoE service with 32K context, tensor-parallel across Bravo+Charlie GPUs, and reasoning token support. On-demand — does not auto-start on reboot.

Vector database for embeddings, semantic retrieval, and RAG storage.

Fast CPU-based text-to-speech using Piper neural voices

Advanced voice cloning and text-to-speech with StyleTTS2

Flow-matching text-to-speech with zero-shot voice cloning

FFMPEG-based audio format conversion and processing

Stable Diffusion image generation with ComfyUI workflows (native systemd service)

Claude Code conversation export to LibreChat format

Semantic search API for AI conversations with vector similarity and hybrid search

Semantic knowledge base for infrastructure docs and lessons learned (744 points, 28 topics, Qdrant + Qwen3-Embed)

Document format conversion for the Haiven ingestion pipeline — converts PDF, DOCX, PPTX, XLSX, HTML, and images to Markdown/JSON using Docling (IBM Research). Primary consumer: haiven-knowledge IngestionRouter.

OpenAI-compatible API gateway with unified model access, virtual keys, and Langfuse observability

Web-based orchestration platform for Claude Code containers with terminal access, MCP configuration, and LLM routing

Autonomous 9-state web research pipeline with LLM synthesis, SearXNG search, Crawl4AI extraction, work-hub task integration, haiven-knowledge auto-ingest, and Langfuse tracing. Supports task_id for artifact write-back.

Personal audiobook recommendation engine with semantic search, weighted scoring, and Libation import

Tri-engine speech-to-text with NVIDIA Canary-1b-v2, Parakeet-TDT-0.6B-v2, Whisper Large v3 Turbo, and pyannote speaker diarization

Automated meeting transcription and note-generation pipeline (7-stage: transcribe → clean → infer metadata → notes → validate → extract → deliver) with v2 edit mode, partial re-runs, and 31 configurable settings

Gemma 4 26B-A4B FP8 MoE inference via vLLM. Vision + tool calling + thinking mode. 256K context with tiny KV cache (5.2 GB at full context). Primary structured output model on Delta GPU.

Gemma 4 E4B BF16 inference via vLLM. Only Gemma 4 model with audio input (30s clips, 16 kHz). Also supports image and video. 128K context.

MedGemma 4B BF16 abliterated medical vision inference via vLLM. SigLIP encoder for radiology, dermatology, histopathology, ophthalmology. 128K context.

AI-powered meeting assistant with SSE streaming chat, reasoning blocks, real-time transcription via haiven-transcribe, speaker diarization, knowledge base search, and chat export

Integration-first workspace — IMAP email, meeting transcription, document import (PDF/DOCX/EML/HTML/CSV), task management, and AI-assisted drafting. 37 endpoints across tasks, meetings, taxonomy, import, audio transcription, email connector, backfill, and webhooks.

Voice-to-content pipeline with 6 content types (Entry, TIL, Link, Quote, Build, Site Page), 2 voice registers (Authority, Conversational), and a spoken feedback loop. Records voice notes, transcribes via haiven-transcribe, drafts via Seed-36B, saves as markdown with YAML frontmatter.

RAGAS evaluation service for the haiven-knowledge RAG pipeline. Measures retrieval quality (Context Precision) against a 25-question golden dataset using GLM-4.7-Flash as judge. Quality gate: >= 0.65.

Cross-encoder reranking service for the Haiven knowledge pipeline. Scores (query, passage) pairs using Qwen3-Reranker-4B-seq-cls via sentence-transformers CrossEncoder on the Delta GPU (port 8460). Internal service — no public domain. OFFLINE / on-demand — stopped 2026-03-30.

Multi-channel notification dispatcher for Haiven agent services. Routes notifications to email (SMTP via Mailpit), ntfy.sh push, or Home Assistant TTS based on a YAML routing table keyed by source_agent.

Scope-aware briefing agent — generates daily summaries, end-of-day reports, task artifact reviews, and knowledge context assembly. Pulls from work-hub and haiven-knowledge, generates via GLM-4.7-Flash, delivers through notification-hub. Scheduled Mon–Fri via systemd timers.

Central AI orchestrator — deployed intent classification via gemma4-26b, session management (Redis), and agent dispatch for 17 intents across briefing, email, scheduling, research, tasks, and knowledge domains

Full-duplex voice pipeline gateway — sequences STT (haiven-transcribe), intent classification (haiven-orchestrator), and TTS (haven-tts-gateway) into a single voice interaction. Includes confirm flow for voice-driven action approval.

AI chat frontend with multi-provider support