So I built a runtime to fix that.

The problem

If you use Claude Code, Copilot, or Codex, you've probably created Agent Skills, those SKILL.md files that tell the AI what to do.

I had a bunch of them. But they were stuck. I couldn't plug them into a product, trigger them from a webhook, or let any service call them with a POST request.

Each skill was trapped inside the tool that created it.

What I wanted

take a SKILL.md → get a POST /run endpoint

Enter fullscreen mode

Exit fullscreen mode

No new framework to learn. No infrastructure to set up. Just point at a skill, configure the model, and deploy.

What I built

Skrun, an open-source runtime that takes Agent Skills and turns them into callable APIs.

skrun init --from-skill ./my-existing-skill

reads SKILL.md, generates agent.yaml

skrun deploy

validates, builds, pushes

→ POST http://localhost:4000/api/agents/dev/my-skill/run`

Enter fullscreen mode

Exit fullscreen mode

Then you call it:

curl -X POST http://localhost:4000/api/agents/dev/code-review/run \  -H "Authorization: Bearer dev-token" \  -H "Content-Type: application/json" \  -d '{"input": {"code": "function add(a,b) { return a + b; }"}}'

curl -X POST http://localhost:4000/api/agents/dev/code-review/run \  -H "Authorization: Bearer dev-token" \  -H "Content-Type: application/json" \  -d '{"input": {"code": "function add(a,b) { return a + b; }"}}'

Enter fullscreen mode

Exit fullscreen mode

{  "status": "completed",  "output": {  "score": 60,  "issues": [  {"severity": "warning", "description": "Use const instead of var"}  ],  "review": "Lacks error handling..."  } }

{  "status": "completed",  "output": {  "score": 60,  "issues": [  {"severity": "warning", "description": "Use const instead of var"}  ],  "review": "Lacks error handling..."  } }

Enter fullscreen mode

Exit fullscreen mode

Multi-model

You pick the provider in agent.yaml, not in your code. Anthropic, OpenAI, Google, Mistral, Groq. If one fails, it falls back to the next.

model:  provider: google  name: gemini-2.5-flash  fallback:  provider: openai  name: gpt-4o

model:  provider: google  name: gemini-2.5-flash  fallback:  provider: openai  name: gpt-4o

Enter fullscreen mode

Exit fullscreen mode

Tool calling

Two approaches.

You can bundle your own CLI tools with the agent. Create a scripts/ directory, write whatever you want (shell, Node, Python), declare them in agent.yaml:

tools:

• name: eslint_check script: scripts/eslint-check.sh description: "Run ESLint on JavaScript code"`

Enter fullscreen mode

Exit fullscreen mode

The LLM calls the tool when it needs to. Skrun executes the script, returns the result. Your agent can run a linter, query a database, call an internal API.

Or use MCP servers. Any MCP server from the npm ecosystem works via npx:

mcp_servers:

• name: browser transport: stdio command: npx args: ["-y", "@playwright/mcp", "--headless"]`

Enter fullscreen mode

Exit fullscreen mode

Stateful

Agents can persist key-value state across runs. Run the same agent twice, it remembers what happened last time.

Roadmap

v0.1 runs on a local registry server. Next up:

Cloud deploy (the architecture has a RuntimeAdapter interface ready for sandboxed VMs), caller-provided API keys, streaming responses, and a hub to discover and share agents.

The numbers

4 packages (@skrun-dev/schema, cli, runtime, api), 10 CLI commands, 154 tests, 6 demo agents, MIT license.

Try it

npm install -g @skrun-dev/cli git clone https://github.com/skrun-dev/skrun.git cd skrun && pnpm install && pnpm build

npm install -g @skrun-dev/cli git clone https://github.com/skrun-dev/skrun.git cd skrun && pnpm install && pnpm build

Enter fullscreen mode

Exit fullscreen mode

Set a Google API key in .env, start the registry (pnpm dev:registry), and follow the "Try an example" section in the README.

github.com/skrun-dev/skrun

I'd love feedback on: The agent.yaml format (does the I/O contract make sense?), the skill import flow, and what agents you'd build with this.