llm-endpoint — call a configured chat/messages API

This skill calls an OpenAI- or Anthropic-compatible HTTP endpoint and returns the reply text. It is a raw model call, not an agent: it sends one prompt and gets one completion. Unlike codex / cursor-agent / opencode, it does not read files, run commands, or explore a repo — you must feed it the context (put it in the prompt, or pipe it in).

Use it when you want a specific model behind a gateway/router/proxy for: a second opinion, a review of a diff/snippet you provide, classification, or text generation — without the overhead (or autonomy) of a full coding-agent CLI.

Setup — secrets stay local, never in git

The base URL and API key live in a local config file outside any repo so they are never committed:

mkdir -p ~/.config/llm-endpoint
cp ~/.claude/skills/llm-endpoint/config.example.sh ~/.config/llm-endpoint/env
chmod 600 ~/.config/llm-endpoint/env
$EDITOR ~/.config/llm-endpoint/env     # fill in BASE_URL, API_KEY, MODEL, KIND

The script reads, in order: existing env vars → $LLM_ENDPOINT_ENV_FILE → ~/.config/llm-endpoint/env. Vars:

The script itself, config.example.sh, and this doc contain no secrets — they are safe to publish. The real ~/.config/llm-endpoint/env is the only place the key lives; keep it out of version control.

Usage

scripts/llm-call.sh (needs curl + jq):

S=~/.claude/skills/llm-endpoint/scripts/llm-call.sh

# Prompt as arg or via stdin; model/kind default to config
"$S" "Explain optimistic locking in two sentences."
printf 'Summarize this repo in one line.' | "$S"

# Pipe context in — e.g. review the current diff
{ echo "Review this diff for bugs, cite the line:"; git diff; } | "$S"

# Pick model / shape explicitly
"$S" --model provider/model --kind messages "your prompt"
"$S" --kind chat --model some-openai-model "your prompt"

# Options
"$S" --system "You are a terse code reviewer." --max-tokens 800 --temperature 0.2 "..."
"$S" --models          # list available model ids
"$S" --raw "..."       # full JSON response instead of just the text

--kind chat → POST {base}/chat/completions (OpenAI shape, Authorization: Bearer). --kind messages → POST {base}/messages (Anthropic shape, x-api-key

• anthropic-version). The script extracts the assistant text from either.

When to use this vs a CLI harness

• Use llm-endpoint when you just need a model's answer over some context you already have (paste/pipe it): second opinions, reviewing a provided diff, Q&A, generation, classification — especially for a model only reachable via your gateway.
• Use codex / cursor-agent / opencode (see those skills) when you need an agent that explores the repo, runs commands, or edits files autonomously. This endpoint can't do that — it only sees what you send it.

In an ultracode workflow

The ultracode-external-agents skill ships an llmCmd() helper so a workflow can delegate a step to this endpoint (the Claude adapter runs the script and relays the answer). Because it's not an agent, gather the context in the command — e.g. pipe git diff in. See that skill's "Raw LLM endpoint backend" section.

Gotchas

• Not an agent. No file/repo access — it only sees the prompt you send. Pipe or paste the context.
• Secrets are local-only. Never hardcode or commit the URL/key; they belong in ~/.config/llm-endpoint/env.
• chat vs messages per model. A gateway may serve a given model on only one shape — if one returns an error, try the other and set LLM_ENDPOINT_KIND. The chat path sends "stream": false because some gateways stream SSE by default (which can't be parsed as a single object — and a streaming failure can even surface as a 502 stream_early_eof that non-streaming avoids). The script still falls back to reassembling SSE if a gateway streams anyway.
• Reasoning/"thinking" models + --max-tokens. Models that emit a hidden thinking block spend the token budget on reasoning first; a low --max-tokens can be consumed entirely by thinking (stop_reason: length), leaving the answer text empty. Raise --max-tokens (default 1024) if a reasoning model returns nothing. The script returns only the final answer text, not the thinking.
• HTTP errors are surfaced. A non-2xx response prints HTTP <code> plus the error body to stderr and exits non-zero (instead of silently returning empty).
• Needs curl and jq.
• Don't leak secrets into prompts — you're sending them to a third-party model.