@getdiff/sdk

Open source, model-agnostic reimplementation of the Claude Agent SDK. Same API, any model.

// Before (Claude Agent SDK — Claude only)
import { query } from "@anthropic-ai/claude-agent-sdk";

// After (OpenAgent SDK — any model)
import { query } from "@getdiff/sdk";

for await (const message of query({
  prompt: "Fix the auth bug",
  options: {
    provider: "anthropic",  // or "openai", "google", "bedrock", "ollama", etc.
    model: "claude-sonnet-4-6",
    cwd: "/workspace",
  },
})) {
  if (message.type === "assistant") {
    // Stream assistant responses
  }
  if (message.type === "result") {
    console.log(`Done in ${message.numTurns} turns`);
  }
}

Why

The Claude Agent SDK is the foundation of the Claude Code product suite — skills, hooks, subagents, channels, and more all compose through its query() / AgentDefinition / streaming interface. But it only works with Claude.

GetDiff SDK gives you the same API with pluggable model providers. Swap one string, keep your code.

Features

• Same API — query(), AgentDefinition, SDKMessage, streaming AsyncIterable
• 11 providers — Anthropic, OpenAI, Google Gemini, AWS Bedrock, OpenRouter, Groq, Together, Fireworks, Ollama, and any OpenAI-compatible endpoint
• 11 built-in tools — Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch, AskUserQuestion, TodoWrite, NotebookEdit
• Subagents — agents config for spawning child agents, recursive delegation
• Permissions — 5 modes (default, acceptEdits, plan, bypassPermissions, dontAsk) + allow/deny lists + canUseTool callback
• Hooks — 11 lifecycle events, blocking/non-blocking, PreToolUse/PostToolUse/SessionStart/Stop
• Sessions — JSONL persistence, list/resume/rename/tag
• MCP servers — Connect to MCP servers via stdio, HTTP, or SSE; auto-discover tools
• Project instructions — Auto-loads CLAUDE.md, AGENTS.md, .cursorrules, .windsurfrules, and more

Install

npm install @getdiff/sdk

Providers

Anthropic (Claude)

export ANTHROPIC_API_KEY=sk-ant-...

const messages = query({
  prompt: "Hello",
  options: { provider: "anthropic", model: "claude-sonnet-4-6" },
});

OpenAI (GPT)

export OPENAI_API_KEY=sk-...

const messages = query({
  prompt: "Hello",
  options: { provider: "openai", model: "gpt-4o" },
});

Tools

All 8 built-in tools are available by default. Restrict with tools:

query({
  prompt: "Read the code",
  options: {
    tools: ["Read", "Glob", "Grep"], // Only these tools
    // tools: "none",                // No tools
  },
});

Tool safety features (inspired by OpenCode):

• Bash: banned commands, safe whitelist, output truncation at 30K chars, timeouts
• Edit: fails on ambiguous matches, requires exact string match
• All tools: permission checks before execution

Subagents

query({
  prompt: "Review the code and fix any issues",
  options: {
    agents: {
      reviewer: {
        description: "Reviews code for issues",
        prompt: "You review code for bugs and style issues.",
        tools: ["Read", "Grep", "Glob"],
        model: "haiku",
      },
      fixer: {
        description: "Fixes code issues",
        prompt: "You fix bugs identified by the reviewer.",
      },
    },
  },
});

Permissions

query({
  prompt: "Refactor the auth module",
  options: {
    permissionMode: "acceptEdits", // Auto-approve file edits
    disallowedTools: ["Bash"],     // Block shell access
    canUseTool: async (tool, input, ctx) => {
      if (tool === "Write" && input.file_path?.toString().includes(".env")) {
        return { behavior: "deny", message: "Cannot modify .env files" };
      }
      return { behavior: "allow" };
    },
  },
});

Hooks

import { createHook, createToolBlocker } from "@getdiff/sdk";

query({
  prompt: "Fix the bug",
  options: {
    hooks: {
      PreToolUse: [
        createToolBlocker("Bash", "Shell access disabled"),
      ],
      PostToolUse: [
        createHook(async (input) => ({
          hookSpecificOutput: {
            additionalContext: "Remember to run tests after editing.",
          },
        })),
      ],
      UserPromptSubmit: [
        createHook(async (input) => {
          if (input.prompt.includes("DROP TABLE")) {
            return { decision: "block", reason: "Dangerous prompt" };
          }
          return {};
        }),
      ],
    },
  },
});

Sessions

import { query, listSessions, renameSession } from "@getdiff/sdk";

// Sessions are auto-persisted
let sessionId;
for await (const msg of query({ prompt: "Start a task", options: { cwd: "/project" } })) {
  if (msg.type === "result") sessionId = msg.sessionId;
}

// Resume later
for await (const msg of query({
  prompt: "Continue where we left off",
  options: { resume: sessionId, cwd: "/project" },
})) { ... }

// List and manage
const sessions = await listSessions({ directory: "/project" });
await renameSession(sessionId, "Auth refactor", "/project");

MCP Servers

query({
  prompt: "Query the staging database",
  options: {
    mcpServers: {
      postgres: {
        command: "npx",
        args: ["-y", "@modelcontextprotocol/server-postgres"],
        env: { DATABASE_URL: "postgresql://..." },
      },
    },
  },
});

MCP tools are namespaced: mcp__postgres__query, mcp__postgres__list_tables, etc.

Project Instructions

The SDK auto-loads project instructions from known locations:

| File | Provider | |------|----------| | .openagent/instructions.md | OpenAgent (highest priority) | | CLAUDE.md | Claude | | AGENTS.md | Codex | | .cursorrules, .cursor/rules/*.mdc | Cursor | | .windsurfrules, .windsurf/rules/*.md | Windsurf | | .github/copilot-instructions.md | Copilot | | .amazonq/rules/*.md | Amazon Q | | CONVENTIONS.md | Aider |

Instructions are merged in priority order and injected as the system prompt (unless you provide an explicit systemPrompt).

License

MIT