server : add Anthropic Messages API support

[noname22]

Summary

This PR adds Anthropic Messages API compatibility to llama-server. The implementation converts Anthropic's format to OpenAI-compatible internal format, reusing existing inference pipeline.

Motivation

• Enables llama.cpp to serve as a local/self-hosted alternative to Anthropic's Claude API
• Allows Claude Code and other Anthropic-compatible clients to work with llama-server

Features Implemented

Endpoints:

• POST /v1/messages - Chat completions with streaming support
• POST /v1/messages/count_tokens - Token counting for prompts

Functionality:

• Streaming with proper Anthropic SSE event types (message_start, content_block_delta, etc.)
• Tool use (function calling) with tool_use/tool_result content blocks
• Vision support with image content blocks (base64 and URL)
• System prompts and multi-turn conversations
• Extended thinking parameter support

Testing

• Tests in test_anthropic_api.py
• Tests cover: basic messages, streaming, tools, vision, token counting, parameters, error handling, content block indices

[noname22]

New PR to allow maintainers to edit.
Old PR here: #17425

[noname22]

The RISCV test is getting

The self-hosted runner lost communication with the server. Verify the machine is running and has a healthy network connection. Anything in your workflow that terminates the runner process, starves it for CPU/Memory, or blocks its network access can cause this error.

I'm guessing it's not related to the PR? Any way to retry?

[ngxson]

This PR can be merge when server CI passes. Other CI are not important.

[ericcurtin]

I stumbled across this as it hit conflicts with my PR. I am curious. What models does this work with? With sufficient hardware is this capable of beating Claude cloud models?

[noname22]

Technically it works with pretty much any model but to get anywhere near Claude Sonnet you'd probably need a large, agentic model like MiniMax M2, Kimi K2, Qwen3 Coder 480B-A35B, etc.

That being said, I've had decent results for simple tasks with Qwen3 Coder 30B-A3B and gpt-oss-20b on a single 4090.

In my very subjective experience, the same models tend to perform a lot better with the Claude Code CLI app than with alternatives such as Open Code or gemini-cli and its clones, like Qwen3-Coder (the cli app).

[ericcurtin]

Interesting... If you want to take a quick peek, I fixed the conflicts here:

#17554

although they weren't major conflicts, it was just moving code from one place to another.