Filesystem-first · Git-native · Provider-agnostic

AI Agent Teams,
Human-like workspaces

Matic runs structured AI agent squads on your codebase using git, markdown, and the CLI tools you already know.

matic — my-project
$ matic init my-project
Created org: my-project/
Scaffolded roles: engineer, reviewer, planner
 
$ matic run "add auth to the API"
→ planner Breaking down task into subtasks...
→ engineer Writing code (claude-code)...
→ reviewer Reviewing diff against spec...
PR #42 ready — 3 files, 127 lines
How it works

A team, not a chatbot.

Matic structures AI work the way engineering teams do — with roles, ownership, and shared memory baked into the filesystem.

01 — ORG

One org, many agents

Your org is a directory. Each role is a subdirectory. Agent state, memory, and history live as markdown files — readable by humans and AI alike.

02 — ROLES

Agents with identity

Engineers, planners, reviewers — each role has a persistent system prompt, task queue, and conversation history. They remember what they worked on.

03 — TASKS

Git is the source of truth

Every agent action is a commit. Task handoffs are PRs. The audit trail is your git log. No proprietary database — just version control you already trust.

0new tools to learn
5+provider adapters
agent roles
1init command
The database is a directory

Markdown is
the schema.

No ORM. No migration files. Agent memory, task queues, and inter-agent messages are markdown files structured by convention — readable in any editor, diffable in any git client.

Human-readable by default

Every file your agents produce is a markdown document. Inspect, edit, or override agent state directly in your editor.

Git handles concurrency

Agents work on branches. Merges surface conflicts. No custom locking mechanism — just the distributed system you already know.

Zero vendor lock-in

Export your entire agent team by zipping a folder. Import it on any machine with Bun installed.

my-project/
├── .matic/                         # org config
│   └── org.md
│
├── roles/
│   ├── engineer/
│   │   ├── system.md              # identity
│   │   ├── memory.md              # long-term
│   │   └── tasks/
│   │       ├── done/
│   │       └── active.md          # current
│   │
│   └── reviewer/
│       ├── system.md
│       └── inbox/                 # queue
│
└── messages/                      # agent IPC
    └── engineer→reviewer.md
Provider adapters

Works with every
model you already use.

Swap providers per role, per task, or per environment. Matic's adapter layer keeps your agent logic identical regardless of what's running under the hood.

CC
Claude Code
Anthropic's agentic coding CLI — the default adapter for most engineering roles.
recommended
GC
Gemini CLI
Google's command-line agent with long context windows and grounding.
stable
CX
Codex
OpenAI's coding agent via the Codex CLI adapter for GPT-4o and o3.
stable
OL
Ollama
Run fully local models — Llama 3, Mistral, Qwen — with zero data egress.
local
HF
HuggingFace
Connect any inference endpoint — open models on serverless or dedicated hardware.
beta
GR
Groq
Ultra-fast inference for latency-sensitive roles like code review and linting.
beta
xAI
xAI Grok
Grok models via the xAI API — competitive on reasoning and context length.
beta
+
Bring your own
Implement the 4-method Provider interface to connect any model or API in minutes.
open
Core concepts

Built to get smarter.

Matic isn't just an agent runner — it ships with a self-improvement loop that lets your team evolve its own prompts and workflows over time.

v0 paradigm

Zero code to start

Declare your org structure in markdown. Matic scaffolds the filesystem, wires up the agents, and connects the provider. No TypeScript until you need it — and when you do, you're extending a clean adapter interface, not wrestling a framework.

# roles/engineer/system.md

You are a senior TypeScript engineer.
You write clean, tested, typed code.
You hand off to reviewer when done.

provider: claude-code
model: claude-sonnet-4-6
MetaMatic

The self-improvement loop

A meta-agent reviews completed tasks, identifies failure patterns, and proposes updates to role prompts — committing them back to git for your review.

Task completes
MetaMatic audits output
Proposes prompt diff
You approve via PR
Roles improve over time

"The best agentic infrastructure is invisible. You shouldn't notice it — you should notice your team shipping faster."

— from the matic.sh design doc

Start with one
command.

No account required. No cloud dependency. Just Bun and a directory.

bunx matic init my-org
Waitlist

Claim early access for
the filesystem-first future.

Join the waitlist for launch access, product notes, and implementation updates. Matic is built for teams that want AI agent systems to feel structured, legible, and operationally real.

Launch accessProduct notesImplementation updates

No spam. Just the important product milestones and the thinking behind them.

launchwaitlist@matic.sh

No spam. Just launch access, product notes, and implementation updates.