Serverless Framework

The Serverless Framework deploys applications to AWS from one serverless.yml: functions, the events that trigger them, and any other AWS resources they need, as one stack per stage.

Start here

• In every project with a serverless.yml, run serverless agent setup in that directory, even when this skill is already installed. It adds the Framework's project skills (one per feature) to the project and reports sign-in, AWS credentials, and the service, with a fix for anything missing. Commit the project skills and keep them even when the service doesn't use a feature yet: each one loads only when a task involves its feature, including adding it.
• Sort out sign-in with the user before the work that needs it: every command that builds or deploys needs it and AWS credentials (references/cli.md), serverless package included.
• Give a database or other data store (DynamoDB, RDS, S3 buckets) its own service, deployed once to a shared stage; the app service reads its names and ARNs through ${param:...} (references/multi-service.md).
• Develop a feature test-first on a personal stage, with serverless dev, unit tests, and integration tests against the deployed stage (references/development-workflow.md).
• Don't answer from memory. serverless agent docs lists the documentation that ships with the installed CLI, matching its version and with no network needed; serverless agent docs <path> prints a page.

First deploy of a service

1. serverless agent setup in the service directory.
2. Sign-in and AWS credentials, resolved with the user.
3. serverless package, which catches config and build errors before anything deploys. In a Compose project, deploy the services this one reads outputs from first (a data service, for example), then package it.
4. serverless deploy --stage ${STAGE:-$USER}: the developer's own stage, not the shared dev (references/package-json-scripts.md). Use $USER only when it is their name. If their name differs, deploy with it and have them set STAGE; if you don't know it, or $USER is generic (root, ubuntu), ask.
5. Verify with a real request to the endpoint, or serverless invoke and serverless logs.

Where to look

• A new service; serverless.yml keys, events, IAM: references/serverless-yml.md.
• A starting point to adapt, or a working service for a runtime, event, or integration: the Serverless Framework examples.
• Extending the Framework past its built-in features (check serverless agent docs first): a plugin from the Serverless plugins, added with serverless plugin install --name <plugin>. The Framework builds JavaScript and TypeScript itself, so it needs no bundler plugin (serverless-esbuild, serverless-webpack, serverless-plugin-typescript), nor the plugins it now includes, such as serverless-python-requirements.
• A database or other stateful resource, several services, values passed between services or stages, Compose, personal stages: references/multi-service.md, with references/package-json-scripts.md.
• Adding or changing a feature (serverless dev, unit and integration tests): references/development-workflow.md.
• Python functions (packaging dependencies, building for Lambda from macOS or Windows, the local loop): references/python.md.
• Operating the CLI, credentials, CI/CD, troubleshooting: references/cli.md.
• The agent commands (setup, docs, skills, inspect, the CLI's own MCP server serverless mcp, dev): references/agent-commands.md.
• Upgrading from v1–v3, or cleaning up older v4 config: the serverless-upgrade skill.
• Running the deployed stage's functions on this machine while developing, by a person or an agent (serverless dev, --detach, dev wait, dev stop): the serverless-dev-mode skill.
• Hosting an MCP server on AWS: the serverless-mcp skill.
• Untrusted, AI-generated, or per-session code in isolated compute: the serverless-sandboxes skill.

agent setup installs these four project skills into the project; where they are not installed, serverless agent skills read <name> prints them.

Gotchas

• serverless with no arguments waits at an interactive template picker. Write serverless.yml yourself; a person at a terminal can run it to scaffold a project.
• A project whose frameworkVersion pins v1–v3, or whose devDependencies list serverless v1–v3, cannot run agent setup or other v4 commands: set frameworkVersion: '4', remove the local copy (npm uninstall serverless), then run serverless agent setup and follow the serverless-upgrade skill.
• serverless package checks the config and the build; checks against the AWS account, such as a resource name already taken, happen at deploy.
• From a tool call, start serverless dev with --detach and --stage <you>, and end it with serverless dev stop, which restores the functions; when a session was killed, serverless dev restore --stage <you> does (the serverless-dev-mode skill).
• When serverless login lists several orgs, ask which one the service belongs to and write org: into serverless.yml (references/cli.md).
• A configuration change pushed with serverless deploy function stays after a later serverless deploy, which changes only what differs from the last template. Run serverless deploy function -f NAME again from the configuration you want.
• An npm warn install-scripts line naming serverless during npm i -g serverless is harmless: the CLI downloads its binary on first run.