Your agents.
Your rules.
Your way.

Obsidian AI is where teams build and run AI agents that actually work — with full control over every model, tool, memory, and channel.

No credit card required · 1,000 free credits on signup

Obsidian AI

live

Real-time streaming chat with live tool call visualization

Sessions

Research session

Code review

Data analysis

What's the weather in Tokyo and summarise today's AI news?

Ask your agent…

Stop prompt-engineering around limitations. Obsidian gives your agents real memory, real tools, and real channels — so they do the work, not just answer questions.

FIG 0.1

Configure, don't code.

Pick a model, wire in tools, attach a knowledge base. Go from zero to a working agent in under five minutes — no infrastructure, no boilerplate, no PhD required.

FIG 0.2

Agents that work together.

Chain agents into teams and pipelines. A coordinator routes queries, specialists handle them, a summarizer delivers the result — while you watch every handoff in real time.

FIG 0.3

Live wherever your users are.

Ship to WhatsApp, a webhook, or a cron schedule with one click. Full streaming traces, per-session memory, and audit logs from day one.

Visual DAG editor

Draw it. Run it.

Design workflows on a canvas. Each node is an agent, a tool call, or an LLM step. Fan out in parallel, branch on conditions, join results — then hit run and watch it execute live.

—Drag-and-drop node editor
—Parallel fan-out and join
—Conditional branching
—Live execution traces per node

Multi-agent teams

One problem, ten agents on it.

Coordinator routes. Specialists execute. Summarizer delivers. Build teams that divide the work, share context, and hand off cleanly — with every decision visible in the trace.

—Coordinator + specialist pattern
—Shared team context
—Real-time handoff traces
—Any LLM per agent role

Knowledge bases

Your data, inside every answer.

Upload docs, connect sources, attach a knowledge base to any agent. Retrieval happens automatically — no prompt stuffing, no manual chunking, no babysitting.

—PDF, DOCX, TXT ingestion
—Vector similarity search
—Per-agent KB scoping
—pgvector / FAISS backend

Long-term memory

Agents that get smarter over time.

Every session extracts structured memories — preferences, decisions, corrections. The next conversation starts where the last one ended. No context lost.

—Model-extracted memory entries
—User-editable memory store
—Bounded context injection
—Cross-session recall

Tools & MCP

Connect anything. Call everything.

HTTP connectors, custom Python tools, MCP servers — your agent calls them natively. Auth, validation, and streaming output are handled. You just define the logic.

—Custom Python tools
—HTTP / REST connectors
—MCP stdio + SSE transports
—Live tool call streaming

Multi-model

Every model. Zero lock-in.

Swap Claude for GPT or Llama without rewriting a prompt. Assign the right model to the right task. Pay only for what you use.

—Claude Haiku / Sonnet / Opus
—GPT-5.4, GPT-5 mini
—Llama 3.3 70B, DeepSeek V3.2, Qwen3
—Bring your own API keys

WhatsApp Business

Your agent answers.
You close deals.

Connect your WhatsApp Business number and assign an Obsidian agent. It reads every message, calls your tools, checks your inventory, and replies in seconds — 24/7, with zero handoff delays.

DMs & group chats

Handle individual and group conversations with full context per chat.

Voice note transcription

Customers send voice notes — the agent transcribes and responds automatically.

Typing presence

Agent shows typing... while generating — feels fully human.

Multi-number channels

Connect multiple WhatsApp numbers, each mapped to a different agent.

Trigger workflows

A WhatsApp message can trigger any Obsidian workflow — orders, bookings, tickets.

QR-code setup

Scan once in-dashboard. Auth persists across restarts — no re-scan.

9:41

Shoe Store Bot

online

Message

Live activity