Imagine a company
as a git repo.
One framework to run your AI-native company. Every agent, skill, trigger, and workflow is code in a single repository — versioned, reviewable, and yours. We call it the Git-First Company Framework.
One repo is the source of truth.
A Kortix Project is a git repository. Agents, skills, commands, triggers, connectors, and memory are just files. Git gives you persistence and 100% version history for free.
Session = sandbox + worktree
Every run is isolated on its own git worktree. Fan out 50 agents in parallel; overlaps merge like git.
Persist via PR
Anything worth keeping is committed to /PERSIST and reviewed before it touches main.
Local == cloud
kortix init / deploy / start. Your local dev runtime is the sandbox runtime.
Define the whole company in code.
kortix.tomldeclares the runtime — sandbox image, cron and webhook triggers, channels, apps, connectors, and required secrets. The.opencodedirectory holds the agents, skills, commands, tools, and models. Change a file, open a PR, ship.
- Engine- and provider-agnostic config
- Reviewable diffs for every change
- Reproducible from a clean clone
Every session is a real, isolated machine.
Each run gets a full Linux sandbox — shell, filesystem, browser, your connected tools — on its own git worktree. Watch the files and terminal live, and run as many in parallel as you need.
- Live files + terminal for every session
- Dozens of agents in parallel, fully isolated
- Resume, fork, or hand off any session
Your infra. Your models. Your engine.
Open and source-available. Self-host it or run it on Kortix cloud — nothing is locked to a vendor.
Self-host
Laptop, a $5 VPS, your VPC, or fully air-gapped. One-line install.
Any engine
Runs on OpenCode today; built to support Claude Code, Codex, and more.
Any provider
Bring your own API keys or subscription, or use Kortix cloud compute.
One secure layer to your whole stack.
Connect once via OAuth, MCP, REST, or Pipedream. Agents reach every tool through a single scoped layer — credentials injected at runtime, never exposed, governed by per-user and per-agent policy.
Any protocol
OAuth, MCP, REST, Pipedream, CLI — 3,000+ integrations out of the box.
Scoped & injected
One token to the sandbox; credentials injected at the network level, never exposed.
Policy-governed
Allow, block, or require approval per connector, per user, per agent.
Governance for humans and agents.
Secrets manager
Encrypted, injected at the network level, never exposed in plaintext.
Users, groups & policies
Every agent, skill, file, and connector scoped per user and per group.
Human-in-the-loop
Approval gates on sensitive actions; persistence requires a reviewed PR.
Own your perimeter
Self-host or single-tenant. Network egress controls. Data never leaves.