Desktop App for Agentic Development
On your terms.
Kaiden runs your coding agents in isolated sandboxes enforced by OpenShell, NVIDIA's open-source runtime for AI agents, and equips them with the models, tools, and assets they need. Repeatable, preconfigured environments to iterate faster. Experiment safely. Works locally or in your enterprise environment. Built for developers, governed by your platform team.
macOS · Linux · Windows — free, open source
Why Kaiden
Running Claude Code on one task, OpenCode on another, something else in the background. Kaiden gives you a single view across every session.
Define a workspace once: pick the model, load the MCP servers, attach your skills and context. Every time you start a session, the agent has exactly what it needs from the first prompt.
Kaiden doesn't replace the agents you use today. It wraps around them so your workspace config, secrets, and tools travel with you.
Core principles
AI agents touch your code, your tools, your credentials. One bad session can cause real damage. Kaiden runs every agent in its own OpenShell sandbox — a container on Podman or a microVM powered by libkrun — with the filesystem, network, and credentials locked down, so the rest of your system stays untouched.
A generic AI tool starts from zero every time. Kaiden connects your agents to the tools, APIs, and knowledge that already exist in your organization, so they actually understand the context they're working in.
Run agents locally when you need speed or privacy. Switch to a remote environment when you need more power. Kaiden handles both without changing how you work.
How it works
Install Kaiden application and follow the flow to configure your coding agents, secrets, connect model providers (local or remote), and import your existing MCP servers and skills. Kaiden is ready to orchestrate your coding agents.
Spin up a coding workspace for your project. You choose the coding agent you want, pick the tools it needs and equip the agent with the skills and MCP servers already in your environment. Run multiple agents in parallel, each in its own sandbox.
Use your coding agent the same way you always have, straight from the CLI. The agent runs inside a sandbox, so the blast radius stays contained. You stay in control of what it can reach. No more worrying about what it might accidentally break.
Capabilities
Each agent session runs in its own OpenShell sandbox — a Podman container, or a libkrun microVM with a kernel of its own when you need a VM boundary. The agent runs non-root, confined to declared paths, with deny-by-default egress.
Use frontier models and coding agents like Claude Code or Codex, or go fully open source with local models and agents like OpenCode or Goose. No lock-in.
Devcontainers let you predefine repeatable environments that automatically equip the agent with the right dev tools inside its sandbox.
Store API keys, tokens, and credentials scoped to workspaces or projects. Secrets are injected at runtime, never exposed directly to the agent.
Tasks that don't need a large model run locally. Heavier workloads route to remote environments. Semantic routing picks the right compute automatically based on what the agent actually needs.
Platform engineers define which models, MCP servers, skills, and context are permitted. Governance is enforced through the workspace config.
Built on OpenShell
OpenShell enforces.
Kaiden is not a new sandbox technology. The runtime underneath is OpenShell, NVIDIA's open-source runtime for safe, private execution of autonomous AI agents — kernel-level enforcement, deny-by-default egress, credentials the agent never sees, and formally verified policy changes.
What Kaiden adds is everything between downloading an app and a running sandbox: the toolchain, the gateway, the image, the policy, and a desktop that keeps every workspace in one place. Features we build — provider management, image tooling, policy primitives — flow back upstream.
- Apache-2.0 open source
- 20k+ GitHub stars
- 4 SDKs Python · TS · Go · Rust
- NVIDIA maintained upstream
- Claude Code · Codex · Goose · OpenCode
- Local, self-hosted or frontier models
- Installs and wires OpenShell, drivers and a local gateway
- Builds a sandbox image from your project’s own image
- Declares mounts, secrets, skills, MCP and network in workspace.json
- Gateway — identity, sandbox lifecycle, policy and the prover
- Supervisor — Landlock, seccomp, egress proxy, credential injection
- The agent never holds a secret and never decides policy
For platform teams
IT controls what developers can run.
Security and Platform teams set the rules once: which models, secrets, tools, and MCP servers are permitted. Then developers can build and run agents without needing to ask for permissions every time. Connect to governance platforms that manage AI Assets.
- Define allowed model providers
- Curate and distribute skills across teams
- Manage MCP servers & tools
- Enforce governance via workspace config
Engineer
Open source
Auditable by design.
Kaiden is built by the team behind Podman Desktop, with over 4 million downloads. Same foundations, same philosophy: open source, extensible, and yours to run however you want.
The sandboxes that protect your agent sessions are enforced by OpenShell and run on Podman containers or libkrun microVMs — the same stacks trusted by millions of developers, now securing your AI agents.
Get started
Need help? Read the docs on GitHub.