Pick your mode
llmwiki runs in two modes that share the same three-layer
pipeline (raw/ → wiki/ → site/) but differ on who calls the LLM:
| Mode A — API | Mode B — Agent | |
|---|---|---|
| How synthesis runs | Python → Anthropic API | Claude Code / Codex CLI → slash command |
| API key needed | Yes (ANTHROPIC_API_KEY) |
No (uses your agent's subscription) |
| Batch + parallel | Yes (native API batching) | No (serial, one turn at a time) |
| Cost model | Pay per token (with prompt cache) | Included in your agent subscription |
| Runs headless? | Yes (cron / CI) | No (needs interactive agent session) |
| Best for | Large corpora, scheduled sync, CI | Exploratory + per-session enrichment |
When to pick which
- You have an Anthropic API key and want to batch-ingest 647 sessions once, then schedule a daily top-up: Mode A.
- You use Claude Code or Codex CLI daily and don't want to pay an extra API bill: Mode B.
- You're evaluating llmwiki locally and just want the dummy / Ollama backend: neither — see Tutorial 08 which works with no agent + no API key.
The two modes share
Everything except synthesis:
- Adapters (
claude_code,codex_cli,cursor,gemini_cli,copilot_chat,obsidian, …) work identically. - The static site, graph viewer, lint rules, backlinks CLI, tag family — all mode-agnostic.
sessions_config.jsonis the same file; onlysynthesis.backenddiffers.
Read next
- API mode — Python CLI with prompt caching + batch.
- Agent mode — slash commands, no API key.
- Upgrade guide — how to switch modes safely.
- Epic: #314 · split LLM Wiki into API mode + Agent mode