❄️ Arcus ADK

A standard, easy-to-use Agent Development Kit for Go.

English · 简体中文

---

Arcus gives Go applications one clean way to talk to large language models. It pairs native, protocol-faithful clients for each provider with a unified chat layer that normalizes requests, responses, streaming, and tool calls — so business code is written once and runs against OpenAI, Anthropic, or DeepSeek without change.

It is built on the database/sql driver model: the core imports no provider package, and applications wire providers in with blank imports. The whole kit is standard-library only — zero third-party dependencies.

cli := chat.New()
_ = cli.Use(adapter.OpenAI, openai.Config{APIKey: key})

msg, _ := cli.Chat(ctx, adapter.Request{
    Provider: adapter.OpenAI,
    Data:     &openai.Request{Model: "gpt-4o", Messages: []openai.Message{openai.UserMessage("hi")}},
})
out, _ := chat.Result(msg)
fmt.Println(out.Text)

✨ Why Arcus

• Native packages, no leaky abstraction. pkg/openai, pkg/anthropic, and pkg/deepseek each speak their wire protocol directly (content blocks, SSE events, reasoning_content, …). Use them standalone, or through the unified layer — your choice, not the framework's.
• One entry point, three providers. pkg/chat routes a request by Provider and hands back a normalized result. Switching models is a config change, not a rewrite.
• Driver-registry wiring. Providers register themselves from init(); you enable them with blank imports. No build tags, no central switch statement, no edits to the core to add a backend.
• Tools that work everywhere. Declare a tool once against a tiny func(ctx, json.RawMessage) (*tool.Result, error) interface. Each driver renders it into that provider's native tool shape — the same tool set drives function-calling on all three.
• Streaming, normalized. A single channel of typed chunks (text, thinking, tool_call, stop, usage, error) regardless of backend.
• Config-file friendly. Provider configs are plain structs with snake_case JSON tags; Use also accepts raw JSON decoded straight from your app config.

📦 Install

go get github.com/arbureva/arcus

Requires Go 1.25+.

🗂 Package layout

Package	Role
pkg/chat	Unified entry point — Client, Chat, ChatStream, Result, chunk helpers. Imports no provider package.
pkg/chat/drivers/{openai,anthropic,deepseek}	Provider bridges. Blank-import to enable; each registers itself from init().
pkg/adapter	Neutral envelopes — Request, MessageAdapter, ChunkMessageAdapter, and the Provider constants.
pkg/openai · pkg/anthropic · pkg/deepseek	Native protocol clients, usable on their own.
pkg/tool	Provider-agnostic tool abstraction — Tool, Func, Reflect, Set, Result.
pkg/agent	The agent loop — Run / RunStream, hooks, sub-agents via AsTool. Imports no provider package.
pkg/agent/transcripts/{openai,anthropic,deepseek}	Native-message transcripts. Blank-import to enable; each registers itself from init().
pkg/toolbox	Progressive disclosure — fold tool groups behind one meta-tool the model opens on demand.
pkg/skill	Skills — instruction packs with optional tools, from code (skill.New) or SKILL.md directories (skill.LoadDir).
pkg/mcp	MCP client (stdio & Streamable HTTP) — remote tools surfaced as ordinary tool.Tool values.
pkg/cli	Command-line programs as tools — cli.Command (argv, allowlisted) and cli.Shell.
pkg/ecode	Shared sentinel errors.

🚀 Quick start

Configure providers

Enable each backend with a blank import, then Use it:

import (
    "github.com/arbureva/arcus/pkg/adapter"
    "github.com/arbureva/arcus/pkg/chat"
    "github.com/arbureva/arcus/pkg/openai"

    _ "github.com/arbureva/arcus/pkg/chat/drivers/openai"   // registers the openai driver
    _ "github.com/arbureva/arcus/pkg/chat/drivers/anthropic"
    _ "github.com/arbureva/arcus/pkg/chat/drivers/deepseek"
)

cli := chat.New()
_ = cli.Use(adapter.OpenAI, openai.Config{APIKey: key, BaseURL: "https://api.openai.com/v1"})

Non-streaming

msg, _ := cli.Chat(ctx, adapter.Request{
    Provider: adapter.OpenAI,
    Data:     &openai.Request{Model: "gpt-4o", Messages: []openai.Message{openai.UserMessage("Introduce yourself.")}},
})
out, _ := chat.Result(msg) // *chat.Completion: Text / Reasoning / ToolCalls / StopReason / Usage / Raw
fmt.Println(out.Text)

Streaming

ch, _ := cli.ChatStream(ctx, adapter.Request{Provider: adapter.OpenAI, Data: req})
for c := range ch {
    switch c.Kind {
    case chat.ChunkText:
        fmt.Print(chat.MustText(&c))
    case chat.ChunkThinking:
        fmt.Print(chat.MustThinking(&c))
    case chat.ChunkUsage:
        fmt.Printf("\nUsage: %d\n", chat.MustUsage(&c).TotalTokens)
    case chat.ChunkError:
        return chat.MustError(&c)
    }
}

Tool calling

Declare a tool once; the same Set is advertised to the model and used to dispatch its calls:

type weatherArgs struct {
    City string `json:"city" desc:"City name, e.g. Shanghai"`
}

tools := tool.NewSet(tool.Func("get_weather", "Get the current weather for a city",
    tool.Reflect(weatherArgs{}),
    func(ctx context.Context, raw json.RawMessage) (*tool.Result, error) {
        var a weatherArgs
        if err := json.Unmarshal(raw, &a); err != nil {
            return tool.Errf("bad arguments: %v", err), nil
        }
        return tool.Textf("It is 24°C and sunny in %s.", a.City), nil
    }))

msg, _ := cli.Chat(ctx, adapter.Request{
    Provider: adapter.OpenAI,
    Data:     &openai.Request{Model: "gpt-4o", Messages: msgs},
    Tools:    tools.RequestTools(),
})
out, _ := chat.Result(msg)
for _, call := range out.ToolCalls {
    res, _ := tools.Invoke(ctx, call.Name, call.Args)
    // feed res.Content back as the provider's tool / tool_result message, then call again
}

chat.Result normalizes tool calls into []chat.ToolCall for every provider, so your dispatch loop is identical across backends. Only the follow-up message reconstruction is provider-shaped (OpenAI/DeepSeek tool messages vs. Anthropic tool_use / tool_result blocks).

Agent

Or skip the manual loop entirely. pkg/agent runs the call → dispatch → append cycle until the model stops asking for tools. Message history lives in a Transcript — a per-provider implementation that stores native messages (so Anthropic thinking blocks, DeepSeek reasoning_content, and tool-call turns all round-trip exactly), registered the same way chat drivers are:

import (
    "github.com/arbureva/arcus/pkg/agent"
    _ "github.com/arbureva/arcus/pkg/agent/transcripts/openai" // like drivers: blank-import to enable
)

bot := agent.New(cli, agent.WithTools(tools), agent.WithMaxSteps(8))

tr, _ := agent.NewTranscript(adapter.OpenAI, &openai.Request{
    Model:    "gpt-4o",
    Messages: []openai.Message{openai.SystemMessage("Be terse.")},
})

tr.User("What's 2+2, and what time is it?")
out, _ := bot.Run(ctx, tr)
fmt.Println(out.Text()) // out.Steps / out.Usage hold the full trace

You always know exactly which native types you injected — the seed request is yours, and tr.Messages() hands the history back in the provider's own message type. agent.RunStream does the same loop over the normalized chunk channel, emitting tool_result chunks between turns.

Multi-agent is one line: agent.AsTool(name, desc, subAgent, seedFn) wraps a specialist agent as a tool.Tool; every delegation gets a fresh transcript from seedFn, so contexts stay isolated.

Progressive disclosure, skills, MCP, CLI

Everything above the loop is still just tool.Tool:

// Fold tool groups behind one meta-tool; the model opens what it needs.
box := toolbox.New().
    Add(clock).                                  // always visible
    Namespace("git", "Read-only git inspection", // folded until opened
        toolbox.Tools(gitTool), toolbox.Instructions("Prefer --stat over full diffs.")).
    AddSkills(skills)                            // each skill folds into its own namespace
session := box.Clone()                           // per-conversation open/closed state
bot := agent.New(cli, agent.WithTools(session))  // *toolbox.Box satisfies agent.Tools

// Skills: instruction packs, from code or from SKILL.md directories.
skills, _ := skill.LoadDir("skills") // each subdir with a SKILL.md becomes one skill

// MCP servers: remote tools indistinguishable from local ones.
srv, _ := mcp.Dial(ctx, mcp.Stdio("npx", "-y", "@modelcontextprotocol/server-filesystem", "."))
defer srv.Close()
remote, _ := srv.ToolSet(ctx) // *tool.Set — plug straight into an agent

// Command-line programs: argv in, stdout/stderr back, no shell in between.
git := cli.Command("git", "Inspect the repo.",
    cli.AllowFirstArg("status", "log", "diff"), cli.Timeout(30*time.Second))

📂 Examples

Runnable examples live under example/:

• example/chat — non-streaming chat, one file per provider.
• example/chat-stream — streaming, one file per provider.
• example/chat-tool — the two-turn tool-calling loop, one file per provider.
• example/agent — an interactive REPL around agent.Run, with hooks narrating every tool round-trip.
• example/agent-multi — multi-agent coordination: a specialist wrapped by agent.AsTool.
• example/mcp — an agent driving the MCP filesystem server over stdio.
• example/skill-toolbox — skills loaded from SKILL.md plus toolbox progressive disclosure.
• example/http — the integration template: arcus inside an HTTP service. An OpenAI-compatible layer (/v1/chat/completions — point Cherry Studio, LobeChat, or any OpenAI client at it; the "model" name selects the service) plus a session-owning API, four services (skills-only, cli-only, mcp-only, multi-agent team), config-file startup, SSE streaming. Start here if you're wiring the SDK into your own backend.

🗺 Roadmap

Arcus aims to be a complete, standard ADK for Go. Every higher-level capability wraps the same tool.Tool interface and chat entry point — adopting them requires no change to code already written against Arcus.

• Native provider clients — OpenAI · Anthropic · DeepSeek
• Unified chat entry point with driver registry
• Streaming with normalized chunks
• Tool calling
• Agent — the tool-calling loop with native-message transcripts, hooks, streaming, and sub-agents (pkg/agent)
• MCP — Model Context Protocol tools as first-class tool.Tool values, stdio & Streamable HTTP (pkg/mcp)
• Skills — instruction packs with tools, from code or SKILL.md directories (pkg/skill), foldable via pkg/toolbox
• CLI — command-line programs as tools (pkg/cli), plus a REPL example for running and inspecting agents
• Structured output helpers
• More providers (Gemini, Qwen, …)

📄 License

Released under the MIT License.

---

English · 简体中文