Developing Genkit Go
GoogleOfficialPublished by Google
- Skill
- Tool integration
- Software development
- Apache-2.0
- Medium risk
AutoGineer summary
Written by AutoGineer from the skill’s own files, not by the author
Genkit Go is an AI SDK that provides a unified interface for integrating generative AI into Go applications. It supports features such as model generation, structured output, streaming, tool calling, and complex agent workflows. The framework is designed for observability and testing, offering a CLI for tracing and a developer UI for inspecting model interactions. It includes experimental support for multi-turn agents, persistent sessions, and generative UI (A2UI) surfaces.
What it does
- Provides a unified interface for interacting with various AI model providers.
- Enables the creation of flows for structured, observable AI logic.
- Supports experimental multi-turn agent development with session persistence.
- Offers middleware for common tasks like retries, fallbacks, and artifact management.
- Includes a CLI tool for capturing traces, debugging, and running flows.
- Facilitates generative UI integration via A2UI middleware.
When to use it
- When building AI features, chatbots, or agents within a Go codebase.
- When you need to integrate multiple AI model providers with a consistent API.
- When you require observability, tracing, and structured testing for AI logic.
- When developing conversational interfaces that require state management or human-in-the-loop capabilities.
Advantages
- Unified API across different model providers.
- Strong observability through integrated tracing and CLI tools.
- Separation of concerns between AI logic and application code via flows.
- Extensible middleware system for common AI patterns.
- Experimental support for advanced agent features like branching and artifacts.
Usage examples
- “How do I define a simple flow in Genkit Go?”
- You use genkit.DefineFlow to wrap your AI logic, which automatically enables tracing and HTTP deployment capabilities.
- “How can I debug my AI model calls in the terminal?”
- Use the Genkit CLI by running 'genkit start -- go run .' to capture traces and inspect model I/O in the Developer UI.
Install
Download the skill as a ZIP:
https://autogineer.com/agent-skills/google/developing-genkit-go/downloadUnzip it into the cross-client skills folder of your project:
mkdir -p .agents/skills && unzip developing-genkit-go.zip -d .agents/skills
Documentation by Google
Genkit Go
Genkit Go is an AI SDK for Go that provides generation, structured output, streaming, tool calling, prompts, and flows with a unified interface across model providers.
Hello World
package main
import (
"context"
"fmt"
"log"
"net/http"
"github.com/firebase/genkit/go/ai"
"github.com/firebase/genkit/go/genkit"
"github.com/firebase/genkit/go/plugins/googlegenai"
"github.com/firebase/genkit/go/plugins/server"
)
func main() {
ctx := context.Background()
g := genkit.Init(ctx, genkit.WithPlugins(&googlegenai.GoogleAI{}))
genkit.DefineFlow(g, "jokeFlow", func(ctx context.Context, topic string) (string, error) {
return genkit.GenerateText(ctx, g,
ai.WithModelName("googleai/gemini-flash-latest"),
ai.WithPrompt("Tell me a joke about %s", topic),
)
})
mux := http.NewServeMux()
for _, f := range genkit.ListFlows(g) {
mux.HandleFunc("POST /"+f.Name(), genkit.Handler(f))
}
log.Fatal(server.Start(ctx, "127.0.0.1:8080", mux))
}Core Features
Load the appropriate reference based on what you need:
| Feature | Reference | When to load |
|---|---|---|
| Initialization | references/getting-started.md | Setting up genkit.Init, plugins, the *Genkit instance pattern |
| Generation | references/generation.md | Generate, GenerateText, GenerateData, streaming, output formats, Embed |
| Prompts | references/prompts.md | DefinePrompt, DefineDataPrompt, .prompt files, schemas |
| Tools | references/tools.md | DefineTool, DefineResumableTool interrupts (Interrupted claim, Restart/Respond, WithResume), ai/tool helpers (tool.Interrupt, tool.Fail, tool.AttachParts) |
| Middleware | references/middleware.md | ai.Middleware, ai.WithUse, Hooks (Generate/Model/Tool), built-ins (Retry, Fallback, ToolApproval, Filesystem, Skills, SoftToolErrors, ContextCompression, Budget) |
| Observability | references/observability.md | OpenTelemetry export (otelgenai), tracing.SetInstrumentation, aix.Observer, reflection API outside dev |
| Flows & HTTP | references/flows-and-http.md | DefineFlow, DefineStreamingFlow, genkit.Handler, HTTP serving |
| Model Providers | references/providers.md | Google AI, Vertex AI, Anthropic, OpenAI-compatible, Ollama setup |
Agents (Experimental)
Genkit Go has an experimental agent API for persistent, multi-turn
conversations (sessions, snapshots, interrupts, branching, background execution).
It is gated: initialize with genkit.Init(ctx, genkit.WithExperimental()) or the
constructors panic. Server constructors come from genkit/exp (aliased
genkitx); types and options from ai/exp (aliased aix); session stores from
ai/exp/localstore.
Agent or flow? If the task is conversational, multi-turn, or described as "an agent", "assistant", or "chatbot", build it with
genkitx.DefineAgentrather than hand-rolling aGenerate+ tools loop in a flow. Reach for a plain flow only for single-shot, stateless generation.
For details see:
Agents: defining/serving an agent, running turns, and client- vs server-managed state (start here).
Sessions & persistence: session stores (
localstore.NewInMemorySessionStore/NewFileSessionStore, PostgreSQL), snapshots, and thesessionstoretestconformance suite.Human-in-the-loop / interrupts: pausing for approval/input and resuming.
Branching: forking a conversation from a snapshot.
Background agents: detaching long-running turns and polling.
Working with state: typed custom session state, streamed as JSON patches.
Artifacts: producing and reading named deliverables.
Multi-agent orchestration: delegating to sub-agents, background delegation,
continue_task, and remote agents (DefineRemoteAgent).Advanced custom agents:
DefineCustomAgentfor full turn control.Deploying agents: serving agents over HTTP with
genkit.Handler.
Generative UI (A2UI)
Genkit Go has an A2UI (Agent-to-UI) plugin
(github.com/firebase/genkit/go/plugins/a2ui/exp, imported as a2uix) that
lets an agent stream interactive UI surfaces (cards, lists, forms, buttons),
not just prose. The whole integration is the &a2uix.Surfaces{} model
middleware added via ai.WithUse
on a generate call or agent inline prompt. Go is server-only: it has no A2UI
client/renderer. Render surfaces with a JS or Dart/Flutter client that talks to
your Go agent over HTTP (the server and clients are wire-compatible).
A2UI: server middleware, options, custom catalogs, and security. For the client (rendering surfaces), refer to the A2UI reference in the Genkit JS or Dart skill.
Genkit CLI (recommended)
genkit start unintrusively wraps any Go program that uses the Genkit library, running it unchanged while capturing traces from every Genkit action so you can prove tools were actually called and inspect model I/O from the terminal, even for headless checks. It forwards stdio, so interactive CLI tools that rely on stdin/stdout work without issues. Running the app directly (go run .) skips trace capture, so you're debugging blind. Check install with genkit --version.
Installation:
curl -sL cli.genkit.dev | bashPrimary pattern (default): prefix genkit start -- to your normal run command. This collects telemetry from any Genkit code your program runs, whether triggered from the dev UI, your own web server/web UI, or a plain script. Starts the Developer UI (usually http://localhost:4000) for running flows, model and agent playground, and browsing traces:
genkit start -- go run .
genkit start --noui -- go run . # same, without the Dev UI (still a persistent server)
genkit start -o -- go run . # also opens the browsergenkit start runs until you stop it with Ctrl+C. That is expected and correct for the common cases: a server your web/mobile app calls, or an interactive CLI you exit yourself. --noui only drops the Dev UI; it is not a one-shot command and will not exit on its own. Do not use genkit start as a blocking step in automated/non-interactive contexts; use flow:run (below) for that.
Non-interactive use (agents/CI): add the global --non-interactive flag before -- so the CLI uses defaults and never blocks on a prompt (e.g. the first-run analytics notice): genkit start --non-interactive -- go run . (works with flow:run too).
Run a flow (flow:run): invoke a specific flow by name from the CLI. Append your run command after -- to spin up the runtime just for this run (the command runs as-is to register your flows):
genkit flow:run myFlow '{"data": "input"}' -- go run .
genkit flow:run myFlow '{"data": "input"}' --stream -- go run . # with streaming
genkit flow:run myFlow '{"data": "input"}' --wait -- go run . # wait for completionThis is self-terminating: it runs the flow once, prints a Trace ID, then exits, so it's the right choice for a quick, non-interactive check (unlike genkit start). Traces for this run can be inspected using the trace commands below.
Debugging with traces: the fastest way to see prompts, model inputs/outputs, tool calls, latencies, and errors. Inspect from the terminal after any run under genkit start:
genkit trace:list # find recent trace IDs
genkit trace:get <traceId> # full trace details (inputs, outputs, tool calls, errors)
genkit trace:get <traceId> --format json # machine-readable JSON, safe to pipe into jq or other parsersFor machine-readable output, pass --format json to get clean JSON you can pipe into jq or other parsers. The default output is human-oriented (banner/log lines, possible truncation on large traces), so don't pipe that form directly; use --format json, grep, or the Dev UI trace viewer.
Documentation:
genkit docs:search "streaming" go
genkit docs:list go
genkit docs:read go/flows.mdSee references/getting-started.md for full CLI and Developer UI details.
Key Guidance
Pass
gexplicitly. The*Genkitinstance returned bygenkit.Initis the central registry. Pass it to all Genkit functions rather than storing it as a global. This is a core pattern throughout the SDK.Wrap AI logic in flows. Flows give you tracing, observability, HTTP deployment via
genkit.Handler, and the ability to test from the Developer UI and CLI. Any generation call worth keeping should live in a flow.Verify with traces, not a blind run. Running the app directly (
go run .) does not capture dev traces. See the Genkit CLI section for how to run your app and capture traces.Use
jsonschema:"description=..."struct tags on output types. The model uses these descriptions to understand what each field should contain. Without them, structured output quality drops significantly.Write good tool descriptions. The model decides which tools to call based on their description string. Vague descriptions lead to missed or incorrect tool calls.
Use
.promptfiles for complex prompts. They separate prompt content from Go code, support Handlebars templating, and can be iterated on without recompilation. Code-defined prompts are better for simple, single-line cases.Use
DefineResumableToolfor interrupts. Claim paused calls with the tool'sInterruptedmethod and resume withai.WithResume.ai.InterruptWith,RestartWith/RespondWith, andWithToolRestartsare deprecated;genkitx.DefineInterruptibleToolandai/exp/toolwere removed in v1.14.0. See references/tools.md.Reach for built-in middleware before writing one.
Retry,Fallback,ToolApproval,Filesystem,Skills, andSoftToolErrorscover the common cross-cutting needs and compose with each other viaai.WithUse. See references/middleware.md. When you do write custom middleware, allocate per-call state in closures captured byNew, and guard anything thatWrapToolmutates because tools may run concurrently.Look up the latest model IDs. Model names change frequently. Check provider documentation for current model IDs rather than relying on hardcoded names. See references/providers.md.
Files
- ATTRIBUTION.mdAdded by AutoGineer858 B
- LICENSELicence file added by AutoGineer11.1 KiB
- SKILL.md11.0 KiB
- references/a2ui.md11.9 KiB
- references/agents-artifacts.md3.9 KiB
- references/agents-background.md5.3 KiB
- references/agents-branching.md3.6 KiB
- references/agents-custom.md6.8 KiB
- references/agents-deployment.md6.6 KiB
- references/agents-human-in-the-loop.md7.3 KiB
- references/agents-multi-agent.md9.3 KiB
- references/agents-sessions.md8.3 KiB
- references/agents-state.md4.6 KiB
- references/agents.md13.0 KiB
- references/flows-and-http.md4.4 KiB
- references/generation.md5.3 KiB
- references/getting-started.md3.4 KiB
- references/middleware.md19.9 KiB
- references/observability.md4.3 KiB
- references/prompts.md7.0 KiB
- references/providers.md3.6 KiB
- references/tools.md8.3 KiB
references/a2ui.md
A2UI (Agent-to-UI) generative UI
The A2UI plugin (github.com/firebase/genkit/go/plugins/a2ui/exp, imported as
a2uix) brings A2UI, a transport-agnostic, JSON-based
streaming UI protocol, to Genkit Go agents. The plugin is in preview: it lives
under the .../a2ui/exp import path (package exp, aliased a2uix following
aix and genkitx) and its APIs may change in any minor version release.
A2UI builds on the experimental agent API
(genkit.Init(ctx, genkit.WithExperimental())). Read Agents
first if you have not.
An A2UI-enabled agent streams more than prose. It streams interactive UI
surfaces (cards, lists, forms, buttons) that a client renders incrementally
as the model responds. The entire Go integration is a single model middleware:
add &a2uix.Surfaces{} to a generate/agent's ai.WithUse and nothing else changes.
Go is server-only: bring your own client
Go has no A2UI client/renderer. The Go plugin is the server half: it injects
catalog capabilities into the prompt and rewrites the model's a2ui blocks into
a2ui data parts on the stream. To actually render surfaces in a browser or app,
use a JS or Dart/Flutter client that talks to your Go agent over HTTP. The Go
server and the JS/Dart client are wire-compatible (same application/a2ui+json
data parts, same catalog ids), so mix and match freely.
For the client half, refer to the A2UI reference in the Genkit JS or Dart
skill (the a2ui.md reference in developing-genkit-js / developing-genkit-dart).
Read their client sections and ignore their server/a2ui() middleware sections
(that is the JS/Dart server, replaced by the Go server below):
JS: browser rendering with an
@a2ui/*renderer (Lit/React/Angular) plus the@genkit-ai/a2ui/clienthelpers (remoteAgent,a2uiEnvelopesFromParts,actionToMessage).Dart/Flutter: rendering with the
genuipackage plus thepackage:genkit_a2ui/client.darthelpers.
The client points remoteAgent at the Go endpoint (for example
remoteAgent({ url: '/api/uiAgent' })), and the client's renderer registers a
catalog under the same catalog id the Go server advertises.
Server: add the a2ui middleware
Add &a2uix.Surfaces{} to the generate call (or agent inline prompt) via
ai.WithUse. That is the whole server-side setup. Registering &a2uix.A2UI{}
in genkit.Init is optional: it only surfaces the middleware and catalogs in
the Dev UI. The middleware works via ai.WithUse regardless.
import (
"context"
"github.com/firebase/genkit/go/ai"
aix "github.com/firebase/genkit/go/ai/exp"
"github.com/firebase/genkit/go/ai/exp/localstore"
"github.com/firebase/genkit/go/genkit"
genkitx "github.com/firebase/genkit/go/genkit/exp"
a2uix "github.com/firebase/genkit/go/plugins/a2ui/exp"
"github.com/firebase/genkit/go/plugins/googlegenai"
)
ctx := context.Background()
// &a2uix.A2UI{} is optional (Dev UI only); the middleware works without it.
g := genkit.Init(ctx,
genkit.WithPlugins(&googlegenai.GoogleAI{}, &a2uix.A2UI{}),
genkit.WithExperimental(),
)
uiAgent := genkitx.DefineAgent(g, "uiAgent",
aix.InlinePrompt{
ai.WithModelName("googleai/gemini-flash-latest"),
ai.WithSystem(
"You help users. Render an A2UI surface whenever a result is " +
"clearer shown than told (weather, comparisons, lists, forms). " +
"Keep prose brief; put the substance in the UI."),
ai.WithUse(&a2uix.Surfaces{}), // defaults to the bundled basic catalog
},
// Server-managed state: the client just passes a session id (remoteAgent
// handles that), history lives in the store.
aix.WithSessionStore(localstore.NewInMemorySessionStore[any]()),
)It works identically on a one-shot genkit.Generate:
resp, err := genkit.Generate(ctx, g,
ai.WithModelName("googleai/gemini-flash-latest"),
ai.WithPrompt("Show me the weather in Tokyo"),
ai.WithUse(&a2uix.Surfaces{}),
)Middleware ordering
A2UI keeps per-turn streaming state (a stream parser and its minted surface ids)
for the model call it wraps. Place any retrying/fallback middleware (which
re-invokes the model) outside A2UI so each attempt gets a fresh A2UI turn.
ai.WithUse(A, B) means A wraps B:
ai.WithUse(
&middleware.Retry{MaxRetries: 5}, // outer: fresh A2UI turn per attempt
&a2uix.Surfaces{}, // inner
)Serve the agent over HTTP
Serve the agent with genkit.Handler (see Deploying agents).
A server-managed agent exposes three endpoints the remoteAgent client expects:
the turn endpoint plus its /getSnapshot and /abort companions.
mux := http.NewServeMux()
mux.HandleFunc("POST /api/uiAgent", genkit.Handler(uiAgent))
mux.HandleFunc("POST /api/uiAgent/getSnapshot", genkit.Handler(uiAgent.GetSnapshotAction()))
mux.HandleFunc("POST /api/uiAgent/waitForSnapshot", genkit.Handler(uiAgent.WaitForSnapshotAction()))
mux.HandleFunc("POST /api/uiAgent/abort", genkit.Handler(uiAgent.AbortAction()))
log.Fatal(server.Start(ctx, "127.0.0.1:8080", mux))If the browser is served from a different origin than the agent (for example a
Vite dev server on :5173), wrap the entire router/mux with permissive CORS
middleware (as shown in Deploying agents)
so that preflight OPTIONS requests are handled automatically without needing to
register duplicate bare paths. Behind a same-origin proxy (for example Vite preview),
CORS is a no-op.
You can also drive the agent directly with curl:
curl -N -X POST 'http://localhost:8080/api/uiAgent?stream=true' \
-H "Content-Type: application/json" \
-d '{"data": {"message": {"role": "user", "content": [{"text": "What is the weather in Tokyo?"}]}}}'Config options
&a2uix.Surfaces{} fields (all optional):
| Field | Default | Description |
|---|---|---|
Catalog | nil | Inline *a2uix.Catalog. Wins over CatalogID, but is not serialized, so prefer CatalogID + LoadCatalog. |
CatalogID | a2uix.DefaultCatalogID ("basic") | Id of a catalog registered with a2uix.LoadCatalog. Resolved from the registry at call time. |
Instructions | a2uix.InstructionsSystem | Where catalog capabilities are injected. a2uix.InstructionsNone injects nothing (supply your own). |
Validate | a2uix.ValidateWarn | ValidateWarn logs and drops bad blocks; ValidateStrict returns an error; ValidateOff skips checking. |
SurfaceID | fresh UUID | Surface-id policy. Set a fixed string to reuse one id per surface; empty mints a fresh UUID per surface. |
Version | a2uix.DefaultVersion ("v0.9") | Protocol version stamped on envelopes. Validated against a2uix.SupportedVersions. |
Use Validate: a2uix.ValidateStrict during development to fail fast on malformed
JSON or components outside the catalog. See Security
for what strict does and does not check.
Custom catalogs
The CatalogID is a catalog id resolved from the Genkit registry. The
bundled basic catalog (a2uix.BasicCatalog(), id a2uix.BasicCatalogID, mirroring
@a2ui/web_core's basic catalog) is the default and needs no registration. A
catalog describes the components the model may emit:
ID: globally-unique URI (also used ascatalogIdoncreateSurface).Components: each ana2uix.CatalogComponentwithName(matches the renderer type),Description(one-line summary), andProps(a compact, model-facing text description of the component's props, kept as plain text to minimize prompt tokens).
Register a catalog with a2uix.LoadCatalog, then reference it by id. Start from
the basic catalog and add your own component:
import a2uix "github.com/firebase/genkit/go/plugins/a2ui/exp"
weatherCatalog := a2uix.BasicCatalog()
weatherCatalog.ID = "https://my-app.org/catalogs/weather.json"
weatherCatalog.Components = append(weatherCatalog.Components, a2uix.CatalogComponent{
Name: "Gauge",
Description: "A circular gauge visualizing a single numeric value.",
Props: "value: number or { path } binding (required); min?: number; max?: number; label?: string; unit?: string.",
})
if err := a2uix.LoadCatalog(g, weatherCatalog); err != nil {
log.Fatalf("loading catalog: %v", err)
}
uiAgent := genkitx.DefineAgent(g, "uiAgent",
aix.InlinePrompt{
ai.WithModelName("googleai/gemini-flash-latest"),
ai.WithUse(&a2uix.Surfaces{
CatalogID: weatherCatalog.ID,
Validate: a2uix.ValidateStrict,
}),
},
aix.WithSessionStore(localstore.NewInMemorySessionStore[any]()),
)You can also load a catalog from a JSON file with
a2uix.LoadCatalogFile(g, "./my-catalog.json") (an object with an id string and
a components array). To surface the bundled basic catalog in the Dev UI
alongside custom ones, call a2uix.RegisterBasicCatalog(g) (optional; the
middleware falls back to it either way). LoadCatalog is idempotent per id.
The client must register a matching renderer under the same catalog id, and
each component Name must match on both sides. Otherwise the model emits a
component the client cannot render. See the A2UI reference in the JS or Dart
skill for how to register a custom component (a genui CatalogItem in Dart, an
@a2ui/* renderer in JS). Catalogs live in the registry under value type a2uix.CatalogValueType
("a2ui-catalog"); the Dev UI lists them at GET /api/values?type=a2ui-catalog.
Security and the trust boundary
Generative UI moves model output into the client UI, so treat every surface an
agent emits as untrusted input. Validate (including ValidateStrict) checks
envelope structure and component type names against the catalog only. It does
not validate component props or data-model values: model-controlled values
such as an Image's url or a Text's inline Markdown pass through untouched.
ValidateStrict is a well-formedness check, not a security boundary.
The client renderer/catalog owns prop sanitization. Whatever renders a surface is responsible for escaping and sanitizing prop values before they reach the UI.
Restrict remote sources at the host. On the web, serve the client with a Content Security Policy that limits
img-srcand other fetch directives to origins you trust.Do not put secrets in the data model. Anything bound into a surface's data model can be echoed back through an action's
context.
For server-side control over props (for example, allow-listing image hosts), add
your own model middleware after &a2uix.Surfaces{} to inspect and rewrite the
emitted a2ui parts.
How it works
A2UI rides on its own part channel: a Genkit data part with mime type
application/a2ui+json (a2uix.A2UIMimeType) whose data is
{ "envelopes": [...] }. On each model call, &a2uix.Surfaces{}:
Sanitizes inbound a2ui parts (a surface action sent back as the next turn, or replayed history) into model-readable text, so a model that does not understand the a2ui mime type can still reason about prior surfaces and user actions.
Injects the catalog's capabilities into the system prompt (unless
Instructions: a2uix.InstructionsNone).Intercepts the model output (streamed chunks and the final message).
Extracts
a2uifenced code blocks from the model's text.Validates them against the catalog (per
Validate).Rewrites them into canonical a2ui data parts.
For a complete, runnable example (a Go backend plus a Vite + Lit web frontend),
see go/samples/basic-middleware/a2ui in the Genkit repo (it serves the same
agent the JS a2ui testapp's web UI talks to).
Requirements and compatibility
Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta
“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.
Security
What our scanner noticed· Medium risk
- Downloads piped into a shell
A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.
Attribution and licence
- Published by
- Publisher
- Google (Official)
- Source
- google/skills · skills/cloud/developing-genkit-go · commit 2194d3b
- Licence
- Apache License 2.0 (redistribution allowed)
- LICENSE (added by AutoGineer from LICENSE)
- ATTRIBUTION.md (written by AutoGineer)
- Synced
AutoGineer is not affiliated with or endorsed by Google or Google.
Download ZIP
The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download developing-genkit-go.zip