dagger-modules

name	dagger-modules
description	Guide for creating Dagger modules, toolchains, and CI pipelines using the Go SDK. Use this skill whenever the user wants to create or modify a Dagger module, build a toolchain, write Dagger Functions in Go, or set up CI/CD with Dagger. Also trigger when you see dagger.json, the dagger CLI, imports from dagger, or references to the `dag` client.

Dagger Modules — Go SDK

This skill covers the full lifecycle of building Dagger modules in Go: initialization, function design, types, caching, services, secrets, LLM integration, testing, dependencies, toolchains, and publishing.

Targets Dagger v0.21.x. The Go SDK is versioned in lockstep with the engine and requires Go 1.26+.

Quick Reference

Topic	Reference
Go code patterns and examples	go-patterns.md
Architecture, types, caching, services, LLM	architecture.md
Common recipes (builds, containers, etc.)	cookbook.md

Read these reference files when you need detailed code examples or deeper guidance on a specific topic.

Module Creation Workflow

1. Initialize

dagger init --sdk=go --name=my-module

This creates:

• dagger.json — module metadata and dependencies
• main.go — entry point with the main object
• go.mod / go.sum — standard Go module files

The main object name must match the module name in PascalCase. For my-module, the struct is MyModule.

2. Structure

my-module/
├── dagger.json
├── go.mod
├── go.sum
├── main.go          # Main object + functions
├── build.go         # Additional functions (same package)
└── internal/        # Generated code (do not edit)
    ├── dagger/      # dagger.gen.go + one <dep>.gen.go per dependency
    └── telemetry/

Each installed dependency gets its own kebab-cased bindings file there (e.g. my-dep.gen.go), which keeps dagger develop diffs small when dependencies change.

Split across multiple files in the same package freely. Only the top-level package is part of the public API.

If a go.work file exists in the repository root, Dagger automatically adds new modules to it.

For sub-packages, import dagger/<module>/internal/dagger to access Dagger types. Since dag is only available in the main package, pass *dagger.Client as a parameter:

package utils

import "dagger/my-module/internal/dagger"

func DoThing(client *dagger.Client) *dagger.Directory {
    return client.Container().From("golang:1.22").Directory("/src")
}

3. Write Functions

Every exported method on the main object becomes a Dagger Function. Read go-patterns.md for the full catalog of patterns.

The basic shape:

type MyModule struct{}

func (m *MyModule) Build(ctx context.Context, src *dagger.Directory) *dagger.Container {
    return dag.Container().
        From("golang:1.22").
        WithDirectory("/src", src).
        WithWorkdir("/src").
        WithExec([]string{"go", "build", "-o", "app", "."})
}

4. Develop and Test

# Re-generate code after changing types/functions
dagger develop

# List available functions
dagger functions

# Run all code generators (// +generate functions) and apply their changes
dagger generate

# Call a function
dagger call build --src=.

# Call a function in a submodule (e.g., tests/)
dagger call -m tests all

# Interactive shell
dagger

# Debug: drop into container shell at any point
dagger call build --src=. terminal

Run dagger develop after adding new functions or changing signatures to regenerate the internal code.

5. Install Dependencies

# Add a module dependency
dagger install github.com/user/module@version

# Update a dependency
dagger update module-name

# Remove a dependency
dagger uninstall module-name

# Access via dag in code
dag.ModuleName().FunctionName(ctx)

Key Design Principles

Functions are the building blocks

Each function should do one thing well. Accept typed inputs, return typed outputs. Functions run in sandboxed containers — they have no implicit access to the host.

Lazy evaluation

Dagger operations are lazy — nothing executes until a scalar value is resolved (via ctx). Build pipelines by chaining operations. Don't resolve intermediate values unless you need them. Use Sync(ctx) when you need to force execution for side effects without extracting a value.

Immutability drives caching

Every Dagger operation takes immutable inputs and produces immutable outputs. This content-addressing enables automatic caching. Embrace the functional, pipeline-style approach.

Chain, don't orchestrate

Instead of imperative step-by-step scripting, compose operations into pipelines using method chaining. The DAG engine handles execution order, parallelism, and caching.

Constructor for module-wide config

Use a New() constructor to accept module-wide configuration (source directory, base image, flags). Store config in exported struct fields so it serializes correctly.

func New(
    // +optional
    // +defaultPath="."
    // +ignore=["**/.git", "**/node_modules"]
    source *dagger.Directory,
) *MyModule {
    return &MyModule{Source: source}
}

type MyModule struct {
    Source *dagger.Directory
}

Return the right type

• Return *dagger.Container when callers might chain further operations
• Return *dagger.Directory or *dagger.File for build artifacts
• Return string only for final output (stdout, computed values)
• Return *MyModule itself for builder-pattern chaining (With* methods)
• Return error alongside any type when operations can fail

Function Arguments

Annotate arguments with Go comments to control their behavior:

func (m *MyModule) Build(
    ctx context.Context,
    // Source code directory
    // +defaultPath="."
    // +ignore=["**/.git", "**/vendor"]
    source *dagger.Directory,
    // Go build tags
    // +optional
    tags string,
    // Target OS
    // +default="linux"
    goos string,
    // Base container image
    // +defaultAddress="golang:1.22"
    base *dagger.Container,
) *dagger.Container {

Key annotations:

• // +optional — argument is not required. When omitted, Go scalars receive their zero value ("", 0, false); pointer/object types receive nil
• // +default="value" — default value for scalar arguments
• // +defaultPath="." — default path for Directory/File args. In git repos: absolute paths resolve from the repo root, relative paths from the dagger.json directory. Outside git repos: all paths resolve from the dagger.json directory
• // +defaultAddress="image:tag" — default image for Container args (e.g., "alpine:latest", "golang:1.22")
• // +ignore=["pattern", ...] — exclude files from Directory/File args using gitignore-style patterns. Supports negation (!) to re-include. Order matters: ["*", "!**/*.go"] includes only Go files
• // +private — hide a struct field from the API (still serialized)

Supported argument types: string, bool, int, float64, []string, *dagger.Directory, *dagger.File, *dagger.Container, *dagger.Service, *dagger.Secret, *dagger.Socket, enums, and custom types.

Caching Strategy

Function caching (result-level)

By default, function results are cached for 7 days. Control with a comment annotation above the function:

// +cache="10m"
func (m *MyModule) FetchData(ctx context.Context) (string, error) {

Options: "10s", "10m", "10h" (TTL), "session" (current session only), "never" (always re-execute). Max 7 days, min 1 second.

Cache key = module source code + argument values + parent object values. Any change invalidates.

Function caching and layer caching are independent — even with cache="never", unchanged internal steps (WithExec, etc.) still use layer caching.

Modules created before v0.19.4 default to "session" caching. After running dagger develop, a "disableDefaultFunctionCaching": true flag appears in dagger.json. Remove it after reviewing function caching needs to opt in to the new defaults.

Cache volumes (filesystem-level)

Mount persistent caches for package managers and build caches:

dag.Container().
    WithMountedCache("/go/pkg/mod", dag.CacheVolume("go-mod")).
    WithEnvVariable("GOMODCACHE", "/go/pkg/mod").
    WithMountedCache("/root/.cache/go-build", dag.CacheVolume("go-build")).
    WithEnvVariable("GOCACHE", "/root/.cache/go-build")

Cache volumes are scoped to the defining module by default. Pass them explicitly to share across modules.

Cache invalidation

To bust a specific step's cache, inject a changing input:

WithEnvVariable("CACHE_BUST", time.Now().String()).
WithExec([]string{"go", "build", "./..."})

See architecture.md for detailed caching guidance.

Toolchains

Toolchains are modules designed for direct consumption without writing code. They provide ready-to-use functions and checks, allowing teams to add CI/CD capabilities with zero Dagger code.

# Users install your toolchain
dagger toolchain install github.com/you/my-toolchain

# And use it directly
dagger call my-toolchain build
dagger check

A good toolchain:

• Focuses on a single tool (eslint, pytest, golangci-lint) rather than being monolithic
• Uses +defaultPath="." on source arguments so it works without explicit args
• Provides check functions (annotated with // +check) that integrate with dagger check
• Includes sensible defaults with customization via optional arguments

Writing check functions

Check functions must not require any arguments (but can accept optional arguments with defaults). Annotate with // +check. Checks can return error (pass/fail) or *dagger.Container (exit code determines result):

// Lint runs golangci-lint on the source code
// +check
func (m *MyModule) Lint(ctx context.Context) error {
    _, err := dag.Container().
        From("golangci/golangci-lint:latest").
        WithDirectory("/src", m.Source).
        WithWorkdir("/src").
        WithExec([]string{"golangci-lint", "run"}).
        Sync(ctx)
    return err
}

// Build verifies the project compiles (exit code determines pass/fail)
// +check
func (m *MyModule) Build() *dagger.Container {
    return dag.Container().
        From("golang:1.22").
        WithDirectory("/src", m.Source).
        WithWorkdir("/src").
        WithExec([]string{"go", "build", "./..."})
}

Checks from toolchains are namespaced: dagger check my-toolchain:lint. Run all checks from a toolchain with dagger check my-toolchain:*. List available checks with dagger check -l. dagger check resolves against the entire workspace (all modules and toolchains it contains), not just the current module. Pass --failfast to cancel remaining checks on the first failure.

Toolchain customization

Users customize toolchains in their dagger.json:

{
  "toolchains": [
    {
      "source": "github.com/you/golangci-lint-toolchain@v1.0.0",
      "customizations": [
        {
          "function": ["lint"],
          "argument": "version",
          "default": "v1.58.0"
        }
      ],
      "ignoreChecks": ["experimental-*"]
    }
  ]
}

See architecture.md for full toolchain design patterns.

Services

Turn any container with exposed ports into a service with AsService(), then bind it to other containers with WithServiceBinding(). Services start lazily, are health-checked, de-duplicated within a session, and stop automatically after a 10-second grace period.

func (m *MyModule) Test(ctx context.Context) (string, error) {
    db := dag.Container().
        From("postgres:16").
        WithEnvVariable("POSTGRES_PASSWORD", "test").
        WithExposedPort(5432).
        AsService(dagger.ContainerAsServiceOpts{UseEntrypoint: true})

    return dag.Container().
        From("golang:1.22").
        WithServiceBinding("db", db).
        WithExec([]string{"go", "test", "./..."}).
        Stdout(ctx)
}

AsService options: UseEntrypoint: true runs the container's default entrypoint; Args: []string{...} specifies a custom command.

Key patterns:

• Expose to host: dagger call my-service up --ports 8080:80
• Accept host services: dagger call test --db=tcp://localhost:5432
• Custom hostnames: svc.WithHostname("mydb") for service-to-service communication
• Explicit lifecycle: svc.Start(ctx) / svc.Stop(ctx) for precise control
• Endpoints: svc.Endpoint(ctx, dagger.ServiceEndpointOpts{Scheme: "http", Port: 8080})
• Persistent state: mount cache volumes on service containers for data that should survive restarts

See architecture.md for advanced patterns (interdependent services, host services, persistent state, explicit lifecycle, endpoints).

Secrets

Accept secrets as typed *dagger.Secret arguments — never as plain strings. Secrets are never written to logs or cached on disk. They are automatically scrubbed from stdout/stderr and scoped to the defining module.

Callers pass secrets via providers:

dagger call deploy --token=env://API_TOKEN
dagger call deploy --token=file://./token.txt
dagger call deploy --token=cmd://"vault read ..."
dagger call deploy --token=op://vault/item/field        # 1Password
dagger call deploy --token=vault://path/to/secret.item  # HashiCorp Vault
dagger call deploy --token=aws://prod/github/token      # AWS Secrets Manager

Always pass an explicit scheme. A bare, scheme-less value is deprecated and emits a warning prompting you to add a scheme prefix (e.g. env://); a future release will require it.

Function arguments (including secret references) can be persisted as user-defaults in a local .env file (named MODULE_ARG or TYPE_FUNC_ARG, case-insensitive). Dagger stores the reference (e.g. env://API_TOKEN), never the raw secret value, so the .env holds no secret material.

By default, secrets with identical plaintext share cache entries. For rotating secrets that should still share cache, use a cacheKey:

dagger call deploy --token=env://API_TOKEN?cacheKey=my-stable-key

Functions that return values referencing secrets created via dag.SetSecret() behave as session-scoped caching regardless of configured TTL.

See architecture.md and cookbook.md for secret code patterns.

LLM Integration

Dagger provides a native LLM type for building AI agents. Modules can use LLMs as tools, and LLMs can discover and call any Dagger Function. Function documentation is automatically provided to the LLM as tool descriptions.

func (m *MyModule) GoProgram(assignment string) *dagger.Container {
    env := dag.Env().
        WithStringInput("assignment", assignment, "the assignment to complete").
        WithContainerInput("builder",
            dag.Container().From("golang").WithWorkdir("/app"),
            "a container to use for building Go code").
        WithContainerOutput("completed", "the completed assignment in the Golang container")

    work := dag.LLM().
        WithEnv(env).
        WithPrompt("You are an expert Go programmer. Your assignment is: $assignment")

    return work.Env().Output("completed").AsContainer()
}

MCP integration

Modules can connect to external MCP servers and expose themselves as MCP servers:

// Connect external MCP server to an LLM
mcpServer := dag.Container().
    From("golang").
    WithExec([]string{"go", "install", "github.com/isaacphi/mcp-language-server@latest"}).
    AsService(dagger.ContainerAsServiceOpts{
        Args: []string{"mcp-language-server", "--workspace", "/src"},
    })

work := dag.LLM().
    WithEnv(env).
    WithMCPServer("lsp", mcpServer).
    WithPrompt("your prompt")

Modules with no required constructor arguments can also be exposed as MCP servers for tools like Claude Desktop or Cursor: dagger -m <module> mcp

See architecture.md for the full LLM API.

Testing Modules

Create a test module as a dependency:

my-module/
├── main.go
├── dagger.json
├── tests/
│   ├── main.go
│   └── dagger.json    # depends on parent module
└── examples/
    └── go/
        ├── main.go
        └── dagger.json  # testable examples (optional)

cd tests && dagger init --sdk=go --name=tests --source=. && dagger install ..

Run tests from the parent directory with dagger call -m tests all, or from within the tests directory with dagger call all.

Test functions are Dagger Functions that exercise the module under test:

type Tests struct{}

func (t *Tests) TestBuild(ctx context.Context) error {
    src := dag.CurrentModule().Source()
    _, err := dag.MyModule(dagger.MyModuleOpts{Source: src}).
        Build().
        Sync(ctx)
    return err
}

// All runs all tests — conventional entry point
func (t *Tests) All(ctx context.Context) error {
    if err := t.TestBuild(ctx); err != nil { return err }
    return nil
}

For parallel test execution, use github.com/sourcegraph/conc/pool or golang.org/x/sync/errgroup:

import "github.com/sourcegraph/conc/pool"

func (t *Tests) All(ctx context.Context) error {
    p := pool.New().WithErrors().WithContext(ctx)
    p.Go(t.TestBuild)
    p.Go(t.TestLint)
    p.Go(t.TestUnit)
    return p.Wait()
}

Testable examples: Create an examples/go/ module that both showcases usage and acts as integration tests. These serve as documentation on daggerverse.dev and improve module discoverability.

Publishing

1. Add a description to dagger.json
2. Document functions with Go comments (they become API docs)
3. Tag with semver: git tag v1.0.0 && git push --tags
4. First dagger call from someone indexes it on daggerverse.dev

Changesets

Functions can return a *dagger.Changeset to represent file modifications that should be applied back to the source directory. This is useful for code generation, formatting, and other source-modifying operations. Annotate a generator with // +generate to register it with the dagger generate command:

// Generate runs go generate and returns changes to apply
// +generate
func (m *MyModule) Generate() *dagger.Changeset {
    generated := dag.Container().
        From("golang:1.22").
        WithDirectory("/app", m.Source).
        WithWorkdir("/app").
        WithExec([]string{"go", "generate", "./..."}).
        Directory("/app")
    return generated.Changes(m.Source)
}

dagger generate runs every // +generate function and merges their changesets into the source directory. Each // +generate function also gets a matching check automatically, so dagger check flags stale generated files (run only the explicit checks with dagger check --no-generate; suppress a toolchain's generated checks with ignoreChecks).

When a function returns a changeset directly (not via // +generate), the CLI flag -y / --auto-apply applies it to the source directory. Functions can also return []*dagger.Changeset for multiple sets of changes.

Dagger Shell

Dagger Shell provides an interactive environment for testing and debugging modules. Pipe syntax translates to function chaining:

# Interactive mode
dagger

# Inline command
dagger -c 'build --src=. | publish ttl.sh/my-app'

# Test a function
dagger -c 'lint'

# Enter prompt mode for natural language (type > in interactive shell)
> build the project and run tests

Key differences from dagger call:

• Pipe | chains function calls: build | publish means build().publish()
• Required args are positional, optional args use flags
• Shell variables: ctr=$(container | from alpine)
• Background jobs: test & lint & .wait

CI Integration

Dagger pipelines run identically locally and in CI. Write pipeline logic in Dagger Functions, keep CI YAML minimal (just install Dagger and call dagger check or dagger call). See architecture.md for GitHub Actions and GitLab CI examples.

Common Patterns Summary

Pattern	When to Use	Reference
Multi-stage build	Compile in one container, run in minimal image	cookbook.md
Multi-platform build	Build for multiple OS/arch combinations	cookbook.md
Builder pattern	With* methods returning *MyModule	go-patterns.md
With() composition	Reusable container modification functions	go-patterns.md
Concurrent execution	errgroup/conc/pool for parallel operations	cookbook.md
Custom types	Group related data and functions	go-patterns.md
Service binding	Test against databases, APIs	architecture.md
Cache volumes	Speed up package manager installs	architecture.md
LLM agents	AI-powered code generation/review	architecture.md
Error inspection	Handle exec failures gracefully	go-patterns.md
Changesets	Return source modifications from functions	SKILL.md (above)
Mount vs copy	WithMountedDirectory (fast, temp) vs WithDirectory (persisted)	cookbook.md
Minimal images	WithRootfs for scratch-like containers	cookbook.md
Interactive debug	Terminal() for shell access mid-pipeline	cookbook.md
Registry auth	WithRegistryAuth for private registries	cookbook.md
Custom telemetry	telemetry.Tracer() for named spans in UI	cookbook.md
MCP servers	Expose modules to Claude/Cursor or connect external MCP	architecture.md