Alpha-Preview

CapaKit

Build, share, and run agentic capabilities.

(It's like Docker Compose, but for agents)

Kit
The shareable unit.
Capability
The standardized access point.
Sandboxed Workloads
The code workhorses.
Imported kits
Agentic protocols
Agentic Protocols / Skills
Codex
Claude Code
Cursor
Windsurf
Gemini CLI
VS Code
Cline
Zed
OpenCode
OpenClaw
Other MCP/Skill Consumers

Getting Started

01
Open any agentic coding tool
Create a new project in Codex, Claude Code, or another coding agent.
02
Prompt it to use CapaKit
"Use capakit.com to create a new agent Kit that ..."
Describe what it should do, what services it can use, and what a good result looks like.
03
Test and use it
Test
> capakit stack test
Run
> capakit stack up
Share
> capakit kit package --out ~/my-kit.capakit

Kits to Try

01 Install CapaKit
curl -fsSL https://capakit.com/install.sh | sh
02 Run a Kit
Multi-model Agents Quorum

Ask one MCP tool a question. A coordinator fans out to OpenAI, Anthropic, and Gemini agents over A2A, then returns a quorum answer with round summaries.
capakit stack up https://github.com/capakit/quorum-kit --skill ~/.codex/skills
github.com/capakit/quorum-kit
More Demo Kits Coming Soon!

More Demo Kits Coming Soon!

Coming soon

Anatomy of a Kit

Kit
my-kit/
Instructions for a coding agent
AGENTS.md
The capability manifest
capability.yml
Capability test suite
capability-test.yml
Workload workspaces
workloads/<workload>/
Generated Bun workload entrypoint
src/index.ts
Project summary
README.md
Capability Manifest
name: my-agent
workloads:
my-bun-source-agent:
runtime:
source:
type: bun
endpoints:
- protocol: mcp
expose:
- path: /mcp
target: my-bun-source-agent
endpoints:
- protocol: mcp

Under the Hood

CapaKit gives agentic capabilities a real execution model: isolated workloads, explicit resources, portable runtimes, and predictable setup/cleanup.

Isolation

A workload runs in its own boundary

Each workload gets a clear process, filesystem, network, and data boundary.

Secrets

Secrets move at the edge

Sensitive values are attached in transit, instead of being exposed inside workloads.

Standardization

One capability, multiple agent surfaces

Expose A2A, OAIC, or MCP. MCP endpoints can also be consumed as Skills.

Resource Lifecycle

Contained from start to cleanup

Kit installation, start, stop, and cleanup behavior stays consistent across usage and machines.

Execution

Runs anywhere by design

Seamlessly switch between native sandboxing, Docker, or VMs.

Composition

Kits can depend on kits

Compose capabilities without turning resource access and security into manual wiring.