[ docs ] // Running models
Harnesses
11434(Ollama's port), so anything already pointed at a local Ollama works with no config change. For the rest, it writes the config for you.arcflare use#
arcflare use opencode qwen3.6-35b-a3bThis loads the model at the best context that fits, writes the harness's config, and launches it. Leave the model out to use the last one. --no-launch writes the config and stops. Arguments after the model go to the harness itself: arcflare use codex -s read-only.
Supported harnesses#
| Harness | How it's wired |
|---|---|
| ArcFlare chat | Built-in streaming REPL. No tools, so it is never asked about approvals. |
| ArcFlare agent | Built in: tools, MCP and skills, no config to write. See the agent. |
| OpenCode | Adds an arcflare provider to opencode.json(c) with the real context limit. |
| Hermes | Uses hermes config set, its own tool, and never hand-edits YAML. Also sets terminal.cwd, otherwise its tools run in your home directory. |
| Hermes Desktop | Same config, launched with hermes desktop --skip-build when the packaged app exists. |
| Codex CLI | [model_providers.arcflare] in ~/.codex/config.toml, with wire_api = "responses". |
Your config is backed up first
Before writing, ArcFlare copies the file to<file>.arcflare-bakand touches only the keys it owns. It refuses to overwrite a config it can't parse, rather than replacing your settings with its own.Auto mode#
The menu asks once whether tools may run without stopping to ask, and remembers the answer. --yolo and --ask override it for one run. Each harness is told in its own language at launch. ArcFlare never writes an approval setting into its config file.
| Harness | Auto mode is |
|---|---|
| ArcFlare agent | No approval prompts. The refusal list for unrecoverable commands still applies. |
| OpenCode | --auto: approves every permission it has not been told to deny |
| Hermes | --yolo: skips prompts, but its hardline blocklist still refuses |
| Hermes Desktop | HERMES_YOLO_MODE=1 in the environment, since no flag exists |
| Codex CLI | --dangerously-bypass-approvals-and-sandbox, because -a never alone still sandboxes |
An approval flag you type yourself always wins, and ArcFlare adds nothing of its own on top of it.
Anything else#
Any OpenAI-compatible client works. Start the server and point the client at it:
arcflare serve| Endpoint | What it is |
|---|---|
| http://127.0.0.1:11434/v1 | Base URL. Any API key is accepted. |
| GET /v1/models | The models the router can serve |
| POST /v1/chat/completions | Chat, streaming and tool calls |
| POST /v1/completions | Plain completions |
| POST /v1/embeddings | Embeddings, for models that produce them |
arcflare ps shows what is loaded, arcflare logs tails the server, and arcflare stop ends it.