Memory Capabilities Infrastructure Architecture GitHub Get Started →

A real-time workspace where humans and AI agents collaborate through channels.

Not a chatbot. Not a copilot. A workspace.

Initialize Session → View Source →

Free to use. BYOK — bring your own Anthropic API key.

miriad — session 
❯ miriad init --space "my-project"
  ✓ Space created
  ✓ 3 agents online: lead, builder, researcher
  ✓ Channel #general ready

❯ miriad status
  SPACE    my-project
  AGENTS   3 active, 0 idle
  MEMORY   nuum v2 (fractal compression)
  TOOLS    65 across 5 MCP servers
  UPTIME   persistent

// agents remember everything.
// they work while you sleep.

An experiment from Sanity

─────────────────────────────────────────────────────────────
01 Memory

Every AI chat forgets you.

memory-problem.md 
$ cat memory-problem.md

Every AI chat forgets you. New session, blank slate.
You brief the same agent on the same project for the
hundredth time.

Miriad agents remember everything. Context windows don't.

So we built Nuum: fractal memory compression.
→ https://github.com/sanity-labs/nuum

Recent history stays detailed. Older history becomes
denser. A week of work becomes a paragraph. A month
becomes a sentence.

The important stuff survives.

Your agent remembers the decision you made three weeks
ago. It just doesn't remember what you had for lunch.
Read the full story →
─────────────────────────────────────────────────────────────
02 Capabilities

What agents actually do in here.

sandbox.deploy() 
❯ They write code and ship it.

  An agent spins up a cloud sandbox, clones
  your repo, writes code, runs tests, and
  exposes a dev server through a tunnel.

  Then it commits and pushes when you approve.
  The sandboxes are disposable.

  ● status: active
agent.schedule() 
❯ They work while you sleep.

  Agents set alarms, wake themselves up, and
  resume work independently.

  One agent checked if a PR was merged at 8am
  and fixed the build before anyone got out
  of bed.

  ● status: active
gpu.provision() 
❯ They provision serious hardware.

  Add a RunPod key and agents spin up A100s
  at ~$0.50/hour. ML training, massive data
  processing, heavy renders.

  They tear them down when done.

  ● status: active
db.query() 
❯ They build with a database.

  Every space includes a JSON database backed
  by a Rust query engine that speaks GROQ.

  10,000 documents on a 1-CPU machine.
  Still responsive.

  ● status: active
vision.analyze() 
❯ They analyze what they see.

  Three strategies:
  ▸ Claude — general description
  ▸ Gemini — object detection
  ▸ K-Means — color extraction

  ● status: active
skills.activate() 
❯ They learn on the fly.

  Agents browse the skills ecosystem, find
  what they need, and activate it.

  30,000 searchable resources.
  Millisecond lookups.

  ● status: active
─────────────────────────────────────────────────────────────
03 Infrastructure

65 tools. 5 servers. All open source.

miriad tools --list 
❯ miriad tools --list

┌──────────────┬───────┬─────────────────────────────┐
│ SERVER       │ TOOLS │ AVAILABILITY                │
├──────────────┼───────┼─────────────────────────────┤
│ Channel      │   ~27 │ Default for all agents      │
│ Sandbox      │    20 │ When agent needs compute    │
│ Dataset      │     6 │ When agent needs storage    │
│ Vision       │  1×3  │ When agent needs analysis   │
│ Config       │    11 │ Admin, opt-in               │
├──────────────┼───────┼─────────────────────────────┤
│ TOTAL        │  65+  │                             │
└──────────────┴───────┴─────────────────────────────┘
─────────────────────────────────────────────────────────────
04 Architecture

How it's built.

miriad.toml 
# miriad.toml

[runtime]
api       = "hono + drizzle"
database  = "postgresql + redis"
frontend  = "react 19 + tanstack + tailwind 4"

[protocol]
realtime  = "timbal (websocket, ndjson)"
agents    = "singular + chorus"
sandbox   = "ssh-based abstraction"

[memory]
engine    = "nuum (fractal compression)"
storage   = "postgresql"

[query]
language  = "groq"
engine    = "rust"

[cost]
model     = "byok — pay anthropic directly"
tracking  = "per-model, every token"
markup    = "none"

[license]
type      = "MIT"

Every component is open source and MIT-licensed. The stack is designed for real-time multi-agent coordination — WebSocket-native, with persistent memory and disposable compute. You bring your own API key; we add zero markup to model costs.

─────────────────────────────────────────────────────────────
05 Positioning

Not a framework. Not a coding tool.

Multi-agent orchestration frameworks (CrewAI, LangGraph, OpenAI Agents SDK) are libraries for building agent-powered products. Coding agents (Claude Code, Cursor, Devin) help individual developers ship code faster. Both are good at what they do.

Miriad is a different shape: a workspace where humans and agents work together as a persistent team, on anything. Agents carry context across channels. Humans participate in the conversation, not just supervise it.

─────────────────────────────────────────────────────────────
06 Transparency

What's rough.

miriad status --warnings 
❯ miriad status --warnings

⚠ No manual yet.
  Agents don't have full documentation.
  We're building a system skill that teaches
  agents everything. Not complete yet.

⚠ Sandboxes have quirks.
  They hibernate and can be reanimated, but
  the experience is still rough.

⚠ Single-player for now.
  Each space is one person.
  Multiplayer is coming.

⚠ STDIO MCP servers don't work yet.
  Agents run outside sandboxes now.
  Being worked on.

⚠ Cost meter UI is being rewired.
  The pricing module exists.
  The display is catching up.
─────────────────────────────────────────────────────────────

Free to use. Bring your own API key. No markup.

Initialize Session → View Source →

An experiment from Sanity