Fantasy

Build AI agents with Go. Multi-provider, multi-model, one API.

1. Choose a model and provider
2. Add some tools
3. Compile to native machine code and let it rip

import "charm.land/fantasy"
import "charm.land/fantasy/providers/openrouter"

// Choose your fave provider.
provider, err := openrouter.New(openrouter.WithAPIKey(myHotKey))
if err != nil {
	fmt.Fprintln(os.Stderr, "Whoops:", err)
	os.Exit(1)
}

ctx := context.Background()

// Pick your fave model.
model, err := provider.LanguageModel(ctx, "moonshotai/kimi-k2")
if err != nil {
	fmt.Fprintln(os.Stderr, "Dang:", err)
	os.Exit(1)
}

// Make your own tools.
cuteDogTool := fantasy.NewAgentTool(
  "cute_dog_tool",
  "Provide up-to-date info on cute dogs.",
  fetchCuteDogInfoFunc,
)

// Equip your agent.
agent := fantasy.NewAgent(
  model,
  fantasy.WithSystemPrompt("You are a moderately helpful, dog-centric assistant."),
  fantasy.WithTools(cuteDogTool),
)

// Put that agent to work!
const prompt = "Find all the cute dogs in Silver Lake, Los Angeles."
result, err := agent.Generate(ctx, fantasy.AgentCall{Prompt: prompt})
if err != nil {
    fmt.Fprintln(os.Stderr, "Oof:", err)
    os.Exit(1)
}
fmt.Println(result.Response.Content.Text())

🍔 For the full implementation and more see the examples directory.

Multi-model? Multi-provider?

Yeah! Fantasy is designed to support a wide variety of providers and models under a single API. While many providers such as Microsoft Azure, Amazon Bedrock, and OpenRouter have dedicated packages in Fantasy, many others work just fine with openaicompat, the generic OpenAI-compatible layer. That said, if you find a provider that’s not compatible and needs special treatment, please let us know in an issue (or open a PR).

Decision models (Jev, Kev)

Beyond chat and embeddings, Fantasy supports evaluation models such as TypeSafe's Jev and the open-weight Kev. These answer typed questions about a state with calibrated probabilities instead of generating text. Providers that support them implement fantasy.EvaluationProvider: providers/vercel (AI Gateway, model typesafe-ai/jev), providers/typesafe (native API, or a running Kev server via WithBaseURL), and providers/kev (Kev fully in-process via llama.cpp, weights downloaded on first use).

ep := provider.(fantasy.EvaluationProvider)
model, _ := ep.EvaluationModel(ctx, vercel.ModelJev)
resp, err := model.Evaluate(ctx, fantasy.EvaluationCall{
    State: "I was charged twice. Fix this today.",
    Questions: map[string]fantasy.EvaluationQuestion{
        "refund":  fantasy.BoolQuestion("Is the customer asking for money back?"),
        "route":   fantasy.ChoiceQuestion("Route this ticket.", map[string]string{"billing": "", "shipping": ""}),
        "urgency": fantasy.ScoreQuestion("How urgent?", "low", "medium", "high"),
    },
})
resp.Answers["refund"].Yes()
resp.Answers["route"].Choice
resp.Answers["urgency"].Level()

Work in Progress

We built Fantasy to power Crush, a hot coding agent for glamourously invincible development. Given that, Fantasy does not yet support things like:

• Image models
• Audio models
• PDF uploads

For things you’d like to see supported, PRs are welcome.

Whatcha think?

We’d love to hear your thoughts on this project. Need help? We gotchu. You can find us on:

• Slack
• Discord
• Twitter
• The Fediverse
• Bluesky

---

Part of Charm.

Charm热爱开源 • Charm loves open source