@lythos/agent-adapter-claude-sdk

Claude SDK adapter for @lythos/agent-adapter. Uses @anthropic-ai/claude-agent-sdk (query() API) instead of the fragile claude -p CLI path.

Install

bun add @lythos/agent-adapter-claude-sdk

Setup

Create .claude-sdk-key in your project root (same directory as package.json):

ANTHROPIC_API_KEY=sk-ant-...
# Optional: custom base URL (e.g. AWS Bedrock, Vertex)
ANTHROPIC_BASE_URL=https://api.anthropic.com

Or set the environment variable directly:

export ANTHROPIC_API_KEY=sk-ant-...

Usage

Import once — the adapter self-registers:

import '@lythos/agent-adapter-claude-sdk'
import { useAgent } from '@lythos/agent-adapter'

const agent = useAgent('claude-sdk')
const result = await agent.spawn({
  cwd: '/tmp',
  brief: 'Write "hello" to output.txt',
  timeoutMs: 60000,
})

console.log(result.stdout)

Why SDK over CLI?

| | claude -p CLI | SDK query() | |---|---|---| | Deferred tool deadlock | ❌ Hangs | ✅ No deadlock | | Stdin pipe bugs | ❌ ARM64 flush issues | ✅ Direct API | | Env pollution | ❌ Nested detection | ✅ Clean | | Tool permissions | --permission-mode bypassPermissions | allowDangerouslySkipPermissions: true | | Output reliability | 6 commits, never produced | ✅ Stable |

How it works

1. Reads .claude-sdk-key on module load (falls back to env vars)
2. Calls query({ prompt, options: { cwd, allowedTools, disallowedTools, permissionMode: 'bypassPermissions' } })
3. Iterates the async generator until type: 'result' message
4. Returns AgentRunResult with stdout, stderr, code, durationMs

License

MIT