Not another AI chatbot.

SyntheticBrew — the open-source
agent brewery

Describe what you need. An AI builder wires up agents, tools, memory, and flows for you.
Self-hosted. Any LLM. Production-ready in one Docker command.

Paused — move cursor away to resume
AI Builder · builder-assistant · glm-5
Ask the builder to configure agents...

Live demo: the AI builder wires up a multi-agent system, then the chat tab runs it. Hover to pause.

Works with OpenAI Anthropic Google AI Groq DeepSeek Ollama + any OpenAI-compatible API

You want AI in your product. Not a 3-month infrastructure project.

Every team building agents runs into the same walls.

Built it yourself

3–6 months of agent plumbing: orchestration, memory, tool calling, retries, observability. Your product waits.

Cloud AI platforms

$500–2,000/mo. Per-token billing that scales with usage. Your data leaves your infrastructure.

Python frameworks

A library, not a product. No REST API, no admin UI, no session persistence. You're still building the runtime.

AI Builder

Agents that build agents

Tell the builder what you need — it configures agents, tools, memory, flows, and gates for you. No YAML to memorize, no SDK to learn.

The builder itself is a SyntheticBrew agent. It runs on the same engine it configures — the best proof the platform is production-ready. Dogfooded end-to-end.

  • One prompt → supervisor + specialists wired up, tools bound, memory scoped
  • Iterate in chat — "add a gate before delivery", "give the researcher memory"
  • Visual canvas stays the source of truth — click any agent to inspect or override
AI Builder assistant panel — describe what you need, agents assemble themselves

One Docker container. Full agent runtime. Your data stays with you.

Your server talks to SyntheticBrew via REST API. Agents think, call tools, stream responses back.

Your Server
Next.js, FastAPI, Go
REST API → ← SSE stream
SyntheticBrew Engine
Docker Container
Any LLM
OpenAI, Gemini, Ollama

Your server handles authentication, then forwards requests to SyntheticBrew.

How it works

From description to production in minutes, not months.

01

Describe

Tell SyntheticBrew what you need

Write what the system should do in plain English. The AI builder proposes agents, tools, and flows — or configure each piece yourself in the visual canvas.

02

Brew

Agents assemble themselves

SyntheticBrew composes the schema: supervisor + specialists, tool bindings, gates between stages, memory scope. ReAct reasoning out of the box.

03

Run

Ship to production

REST API + SSE streaming. Built-in web client and embeddable widget. Memory persists, flows coordinate, Inspect dashboard shows every step.

Every piece ships in the box

Admin Dashboard, AI Builder, Widget generator, Knowledge base — built-in, not bolted on. Use as-is or build your own UI on top of the REST API.

Admin Dashboard

Schemas, agents, MCP servers, models, triggers, audit log — every knob of the engine in one place.

Admin Dashboard — schemas list with AI Assistant panel

Widget Generator

Pick a schema, customize the look, copy one <script> tag onto your site. Chat is live.

Widget Snippet Generator — embed snippet ready to copy

Knowledge Base

Upload PDFs, DOCX, URLs. Agents search via vector similarity automatically — no RAG plumbing on your side.

Knowledge Bases — upload and manage documents for RAG

Agent Detail

Drill into any agent: system prompt, model, tools, spawn rules, memory config. Export/import as YAML.

Agent detail — model, system prompt, tools, spawn rules

Everything your agents need

Built-in, not bolted on. Every capability included in one container.

AI Builder

Describe what you need in plain English. The builder configures agents, tools, flows, and memory — itself running on SyntheticBrew.

ReAct Reasoning

Agents think step-by-step: Reason → Act → Observe → Repeat. Not scripted flows — genuine reasoning.

Memory

Per-schema, cross-session persistence. Agents remember customers, context, and decisions across conversations.

Multi-Agent Flows

Agents coordinate via flow edges, transfer, and spawn. Build teams of specialized agents with gates and loops.

Knowledge / RAG

Upload PDF, DOCX, URLs. Agents search knowledge automatically. Per-schema isolation.

Knowledge Graphs

Declare your domain as typed entities — categories, attributes, relationships. The engine auto-generates MCP tools for deterministic retrieval. No hallucinated IDs, full recall on structured queries.

Embeddable Widget

Generate a chat widget and embed one <script> tag on any site. Connected to your agents, your data.

MCP Tool Ecosystem

Connect to any external service via MCP. Curated catalog, one-click install. Stdio, SSE, Docker transport.

Inspect Dashboard

Full session trace: every reasoning step, tool call, memory access, and decision — searchable.

Recovery & Resilience

Heartbeat monitoring, MCP timeout handling, dead letter queues, circuit breakers for external services.

Updated in 1.4

Stop trusting your LLM to model your domain

Knowledge Graphs let you declare your domain ontology once. The engine auto-generates MCP tools for deterministic entity retrieval — no hallucinated IDs, full recall on structured queries, GitOps-native.

1.4 update: batch fetch, summary projections, server-side sort, and range filters cut typical agent retrieval token cost by ~12×. Split-by-directory bundle layout for catalogs that have outgrown a single file.

You write

JSON Schema

category, brand,
product_attribute

$ brewctl kg apply

Engine

validates refs,
builds tools

Auto-generated

MCP tools

list_brand,
get_category, ...

Result

Agent

grounded, cited,
zero hallucinated IDs

Declarative-first, not extracted

Other platforms ask an LLM to extract entities from your docs. We ask you to declare them. JSON Schema in, typed MCP tools out. Zero extraction drift.

Auto-generated MCP tools per entity

Each entity type produces list_X, get_X, list_X_ids tools. Filterable by any field you mark x-index. Total count returned, so the agent never misses an approved record.

GitOps-native bundles

Your domain lives in git. brewctl kg apply ./my-bundle deploys atomically — all entities valid, or rollback. PR review, branching, rollback, Helm chart integration.

Coexists with vector RAG

Use Knowledge Bases for narrative search and Knowledge Graphs for structured retrieval on the same agent. Memory, Knowledge, Knowledge Graphs — three capabilities, three kinds of memory.

1

Your taxonomy

human-readable, what you already have

category
  └─ footwear  (Footwear)
       ├─ brand: north-aurora
       │         tier: premium
       │         carries: hiking boots
       ├─ brand: stride-co
       │         tier: mid
       │         carries: running shoes
       └─ brand: budget-basics
                 tier: budget
                 carries: casual sneakers
2

Your declaration

JSON Schema in — typed MCP tools out

# schemas/brand.schema.json
{
  "x-id-field": "code",
  "properties": {
    "code":     {"type":"string", "x-index": true},
    "category": {"x-ref": "category", "x-index": true},
    "tier":     {"enum":["budget","mid","premium"],
                 "x-index": true}
  }
}

# Apply atomically:
$ brewctl kg apply ./my-bundle

# Engine now exposes auto-generated tools:
list_brand(filters={category, tier})
get_brand(id)
3

Your agent's answer

grounded, cited, zero invented IDs

User: "premium footwear brands?"

Agent tool call:
  list_brand(filters={
    category: "footwear",
    tier: "premium"
  })

Response:
  1 premium footwear brand:
  • north-aurora

Total returned: 1 (no missed records)

Secure by default

Production-grade defaults you don't have to think about.

Tamper-proof auth

Every API request is cryptographically signed. Forged tokens, downgrade attacks, replay attempts — all rejected at the door.

Zero secret management

No passwords, no rotating shared keys, no .env to leak. Self-hosted runs out of the box. Plug in your existing IdP with one public key.

Bring your own LLM keys

Pass per-customer LLM API keys via request headers — used once, never stored, never logged. Clean per-tenant billing without vendor lock-in.

No double-billing, ever

If usage tracking can't reach the cloud, requests stop — instead of silently double-counting later. Your invoice always matches reality.

Built for real operations

Not demos. Production workloads running 24/7.

Support agent in 5 minutes

Upload your docs to Knowledge, generate a widget, embed one <script> tag on your site. Agent remembers each customer across sessions.

KnowledgeWidgetMemory

AI-first product without the agent team

Embed agents into your SaaS via REST API. Self-hosted — data stays on your infrastructure. No vendor lock-in.

REST APISelf-hostedBYOK

Autonomous data pipelines

Cron triggers start analysis agents. Sub-agents parallelize work. Gates validate output before delivery.

TriggersFlowsGates

How SyntheticBrew compares

Every approach has trade-offs. Here's where SyntheticBrew fits.

Traditional Approach The Problem SyntheticBrew
Cloud AI platforms Per-token pricing, data leaves your servers, locked to one provider Self-hosted with your own API keys. Pay only your LLM provider — no markup
Agent SDKs / frameworks A library, not a product. No API server, no admin UI, no scheduling Complete runtime: REST API, admin dashboard, cron triggers, session management
Visual AI builders Simple chatbots only. No autonomous reasoning, no tool calling, no sub-agents Multi-step reasoning agents that delegate, call tools, and coordinate
Single-model APIs One provider, no orchestration, no memory, no background jobs Mix any models across agents. Built-in RAG, sessions, triggers
Custom in-house build 3–6 months to build, ongoing maintenance, team distracted from product Production-ready in 5 minutes. We maintain the engine — you ship your product

Get started in 30 seconds

No PostgreSQL? No problem — it's included.

$ curl -fsSL https://syntheticbrew.ai/releases/docker-compose.yml -o docker-compose.yml && docker compose up -d

What happens next

  1. Open localhost:8443 — Admin Dashboard
  2. Add your OpenAI / Gemini / Claude API key
  3. Describe what you need in the AI Builder — agents assemble themselves
  4. Open localhost:8443/chat/ to test instantly
  5. Connect your way — see options below

Total time: under 5 minutes. No config files.

Widget

Embed on any site

Generate a chat widget in the Admin Dashboard. One <script> tag — done.

Widget docs →

REST API + SSE

Connect your app

Call POST /api/v1/agents/{name}/chat from your backend or frontend. Streaming responses via SSE.

API reference →

MCP Tools

Extend with custom tools

Write your own MCP server in any language. Agents call it automatically via stdio, SSE, or Docker.

MCP guide →

Powering AI agents in production.

Teams ship real AI products on SyntheticBrew — self-hosted, on their own infrastructure.

Dogfooded
Our own AI Builder runs on SyntheticBrew
Live examples
Open examples

Open source. Community driven.

SyntheticBrew is BSL 1.1 licensed. Free to self-host, embed in your products, and modify.
Converts to Apache 2.0 after 4 years.

Start building agents today

Two ways to get started. Both free.

Public Beta

Cloud

No server to run. Start in minutes.

  • Managed infrastructure
  • Bring your own API key (BYOK)
  • 1 schema · 1,000 agent steps/month
  • 50 MB knowledge storage

Free during Public Beta. Terms may change.

Try Cloud Free →
Community Edition

Self-host

Your infrastructure. Full control.

  • Unlimited schemas, agents & steps
  • Your own API keys, any LLM
  • File & shell tools included
  • Free forever — BSL 1.1 → Apache 2.0
docker compose up -d
Get started →