Covers newer features and developer-facing SDK capabilities.
The ultrathink keyword:
"ultrathink: deeply analyze the pros and cons of this architecture decision"
When you include the ultrathink keyword, Claude allocates up to 32k tokens for internal reasoning before answering. This is useful for complex architecture decisions or tricky debugging.
MAX_THINKING_TOKENS setting:
{
"thinking": {
"maxTokens": 5000
}
}
If this setting is present, it takes priority over the ultrathink keyword.
Using Extended Thinking in the API:
thinking: { type: "enabled", budget_tokens: 5000 }
Claude displays step-by-step reasoning in the thinking block.
Language Server Protocol (LSP) gives Claude IDE-grade code intelligence.
What LSP provides:
Practical impact:
"Refactor this function"
# Without LSP: Claude edits code, and you only discover errors after running
# With LSP: Claude detects type errors as it edits and fixes them immediately
You can use Claude Code's agent loop, tools, and context management via an SDK.
Build an agent in 10 lines:
import { query } from '@anthropic-ai/claude-agent-sdk';
for await (const msg of query({
prompt: "Generate markdown API docs for all public functions in src/",
options: {
allowedTools: ["Read", "Write", "Glob"],
permissionMode: "acceptEdits"
}
})) {
if (msg.type === 'result') console.log("Docs generated:", msg.result);
}
Use cases:
Share team settings:
.claude/team-settings.json:
{
"permissions": {
"allow": ["Read(src/)", "Write(src/)", "Bash(npm test)"]
},
"hooks": {
"PreToolUse": {
"command": "bash",
"args": ["-c", "echo 'Team hook: validating...'"]
}
},
"mcpServers": {
"company-db": {
"command": "npx",
"args": ["@company/db-mcp"]
}
}
}
Shared workflow:
Commit the .claude/ folder to your team repository so everyone uses the same configuration.