Model providers

⚙ Settings → Model provider.

Point a session at something other than your Anthropic account. It stays a real, full claude session with this project's usual agents, skills, hooks, MCP servers, slash commands, output styles, CLAUDE.md and /rewind checkpoints — only the model endpoint changes.

That is not a trick. Everything in the list above is client-side orchestration that never touches the model API, so it survives a backend swap untouched, and Anthropic publishes the wire contract a backend has to meet at code.claude.com/docs/en/llm-gateway-protocol.

The one requirement

The backend must speak POST /v1/messages — the Anthropic Messages API — and it must stream. A gateway that buffers a whole response before relaying it stalls the client.

Backends that do, today:

Backend	Notes
Ollama ≥ 0.14	Native /v1/messages. No prompt caching and no tool_choice, which can make Claude Code pick the wrong tool and loop. Raise num_ctx — see below.
vLLM	Native /v1/messages and count_tokens. Does not return reasoning tokens yet.
llama.cpp (llama-server)	Native /v1/messages. Needs --jinja for tool use.
OpenRouter	Its Anthropic-format endpoint. Tool use and caching vary by model.
OmniRoute	See Plan → Execute & OmniRoute.
A remote/self-hosted box	Any of the above over the network. Use a token.

OpenAI-shaped backends

An endpoint that only serves /v1/chat/completions — LM Studio, most bare local servers, anything "OpenAI-compatible" — is reachable through archeus's own translating gateway. Turn it on in the same settings card: set the gateway target to the OpenAI-shaped URL and archeus runs a loopback proxy that speaks Anthropic Messages to claude and OpenAI Chat Completions upstream.

What the translation cannot carry, and does not pretend to:

• cache_control has no equivalent. It is dropped and logged once per run, so you find out from a line in the proxy window rather than from a cost change you cannot explain.
• Thinking blocks are dropped from request history unconditionally — including on a session you resume after switching backends, which is the case that otherwise produces Invalid signature in thinking block long after the swap and looks unrelated to it.
• count_tokens is answered with the same chars//4 estimate archeus uses elsewhere. There is no dependency-free exact counter that works across providers.

Two operational notes that are not optional:

• On vLLM, tool calling needs --enable-auto-tool-choice --tool-call-parser <parser> server-side. "OpenAI-compatible" guarantees the envelope, never the tool behaviour, and without those flags the model answers in prose where a tool call was required.
• Claude Code aborts a stream that goes quiet for ~90s. The gateway sends keepalive pings for the whole stream, not just until the first token, because a local model doing grammar-constrained tool-call decoding goes silent in the middle of a turn.

Setup

1. ⚙ Settings → Model provider → Backend:
2. Anthropic (direct) — the default. Nothing is overridden.
3. Anthropic-shaped server — a server you already run.
4. OmniRoute — archeus starts the daemon for you.
5. Base URL — e.g. http://localhost:11434 for Ollama.
6. API key — whatever that backend wants, or blank. Write-only: it is never sent back to the browser.
7. Context window — optional, and not probed: most /v1/models responses omit it, and a fabricated number is worse than none. Filling it in turns on a pre-launch warning.
8. Pick the model at launch. On OmniRoute you get a live menu; on a generic server there is no catalogue to read, so you type the model id.

archeus checks the backend is reachable before the session opens. A launch that succeeds and only dies once claude tries the model leaves you in a new console with no path back to the setting that was wrong.

archeus's own calls

Separately from your sessions, archeus makes its own headless claude -p calls — memory extraction, lesson distillation, code review, and the CLAUDE.md / agent / skill / hook / system-prompt generators. Those keep running on Anthropic even when this card points everywhere else, because they run unattended (from a hook, from background threads) and quietly moving them to another model — and another bill — is a surprise rather than a feature.

Run archeus's own calls here too switches them over. They then ask for the model id above rather than the economy model, because that is the only id the backend can resolve. If the backend is unreachable the call fails and says so; it does not fall back to the account you routed away from.

These are the cheapest calls archeus makes and the best fit for a local model: short prompts, no tool use, structured output. Extraction quality does drop on a small model — the memory graph is only as good as what read the transcript.

What does not survive the swap

Stated plainly, because a feature page that only lists what works is not useful. None of these are archeus bugs — they are properties of Claude Code's client or of Anthropic's own infrastructure.

	What happens	What archeus does
Subagents	The Task tool's model parameter is a hard-coded sonnet/opus/haiku enum, and built-in agents carry literal model ids that 401 through any non-Anthropic endpoint.	Strips the model: field from the agent files it syncs, so your project's agents inherit the session's model. Claude Code's built-in agents cannot be fixed from outside.
Prompt caching	Absent on Ollama entirely; varies elsewhere. Every turn bills as uncached.	Nothing to do — cache_control is emitted by claude itself.
Extended thinking	A thinking block carries a signature that must round-trip byte-for-byte to the infrastructure that minted it. A backend that did not mint it cannot produce one, and Claude Code sends the field unconditionally, so an upstream that does not know it answers 400 — failing the whole turn, not just the thinking.	Sets CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 automatically.
web_search	Runs on Anthropic's own servers. Cannot work through a third party at all.	Use an MCP search server instead.
MCP tool search	Claude Code disables it on any non-first-party base URL.	Opt-in toggle — re-enabling only works if your upstream forwards tool_reference blocks.
Cost tracking	There are no published rates for a routed model.	Shows n/a rather than $0.00 — a routed model is not necessarily free.
Remote Control	Disabled by Claude Code on a non-first-party base URL.	—

Local models: the part that is about the model, not the harness

Claude Code's own system prompt is 10k+ tokens before any of your project context. Ollama's default context is 4096. That is already over budget on an empty repo, which is why the pre-launch warning counts the system prompt rather than just weighing your CLAUDE.md.

• Raise the context window (num_ctx on Ollama). This matters more than model size.
• Only a few model families tool-call reliably enough for agentic work — Qwen3-Coder 30B+, GLM-4.7, Llama 3.3 70B, Devstral Small 2, gpt-oss. Below ~7B, or on a model that was never tool-tuned, malformed tool calls are the norm regardless of harness.
• Q4_K_M is the floor for tool calling. Q3 and below breaks the tool-call format before it visibly hurts chat quality, which makes it look like a archeus bug when it is not.
• On vLLM, tool calling needs --enable-auto-tool-choice --tool-call-parser server-side. "OpenAI-compatible" guarantees the envelope, never the tool behaviour.

A note on what this is

Anthropic documents the gateway protocol but states it does not support routing Claude Code to non-Claude models through one. Nothing in the client enforces that; treat this feature as unsupported by Anthropic, not endorsed by them.

Separately, and this one is not ambiguous: archeus never sends your Claude subscription credentials anywhere but Anthropic. The provider key is one you supply for this purpose, stored in archeus's own settings. Reusing Claude Code subscription OAuth tokens in another tool is a Terms of Service violation that Anthropic has actively enforced against other projects — archeus does not do it, and will not gain a setting that does.

Rotation does not apply here

Automatic account rotation moves work to another Anthropic account when the one in use fills its window. A session routed at a provider is not spending an Anthropic account's quota at all, so it is left alone — switching it would replace the whole environment, ANTHROPIC_BASE_URL included, and the call would land silently back on Anthropic.

The same reasoning applies to archeus's own Claude calls: quota.preflight returns early for anything carrying a base URL. If you want a provider-routed session to survive a dead model, that is a different mechanism — the failover proxy, which retries the next model in the profile's list rather than the next account.