Claude Code

Anthropic. Official docs describe LLM gateways. llama.cpp documents an Anthropic-compatible /v1/messages. We have not measured this pair on the rig.

Official CLI speaks the Anthropic Messages API via ANTHROPIC_BASE_URL — not a drop-in OpenAI /v1/chat/completions client.

What it’s for

Anthropic’s coding CLI. Local attach is “where supported”: a gateway or a server that implements /v1/messages.

Local endpoint

llama-server base: http://127.0.0.1:8080
Strata base: http://127.0.0.1:8083
Model name: claude --model <server id> (not documented against our alias)
API key: ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY — any non-empty value if the server ignores it

Tool calling

Official tool use is Anthropic-shaped. Whether llama.cpp’s Messages endpoint maps every Claude Code tool is not documented here.

Long context

Not documented for a local llama-server pairing.

ROCm / Radeon gotchas

Claude Code never talks HIP. Do not set ANTHROPIC_BASE_URL to …/v1 — the client appends /v1/messages itself.

Hybrid local + cloud

Default is Anthropic cloud. Local is a gateway override.

This is the same question as the calculator’sfully localversushybridtoggle. This page does not quote cloud prices.

Official links

Sources

Copy-paste

More on the connect page. Prove http://127.0.0.1:8080/v1/models first.

Claude Code → llama-server :8080

# Official path is Anthropic Messages. The client appends /v1/messages — do not put /v1 on the base.
# llama.cpp documents /v1/messages. We have not measured this pair on the rig.
export ANTHROPIC_BASE_URL=http://127.0.0.1:8080
export ANTHROPIC_AUTH_TOKEN=local-not-a-secret
claude --model qwen

Exact llama-server flags for our daily driver stay TODO in stacks.yaml. This is the port only. Local attach is “where supported.” Completeness of Claude Code tools against llama.cpp Messages is not documented here.

Claude Code → Strata :8083

# Official path is Anthropic Messages. The client appends /v1/messages — do not put /v1 on the base.
# llama.cpp documents /v1/messages. We have not measured this pair on the rig.
export ANTHROPIC_BASE_URL=http://127.0.0.1:8083
export ANTHROPIC_AUTH_TOKEN=local-not-a-secret
claude --model qwen

Exact Strata flags for our rig stay TODO in stacks.yaml. This is the port only. Local attach is “where supported.” Completeness of Claude Code tools against llama.cpp Messages is not documented here.