Unsloth server-side tools versus agent passthrough
Parent: Mac local LLMs: Runtime selection and frontends · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
Unsloth's server-side tools are Python execution, web search and a terminal (bash) tool, and results stream back as `tool_result` events.
These notes link each claim to its source. A source may be a research report hosted on this site rather than the primary document. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- Unsloth's server-side tools are Python execution, web search and a terminal (bash) tool, and results stream back as `tool_result` events. [source]
- A request opts in with `enable_tools: true` and `enabled_tools: ["python", "web_search", "terminal"]` (plus an optional `session_id`) on either the OpenAI-style or Anthropic-style endpoint. [source]
- The API page's troubleshooting says that for built-in tools both `enable_tools: true` and an `enabled_tools` list are required, or the tools do not run. [source]
- The stated reason for the network default is that a leaked API key on a network-exposed server would mean arbitrary code execution on the host. [source]
- With server-side tools off, the model still sees client-side tools sent in `tools` and `tool_choice` in native OpenAI or Anthropic form, and Claude Code, opencode, Cursor, Continue and Cline are listed as working with them. [source]
- Unsloth's Codex guide gives the same instruction as the Claude Code guide: add `--disable-tools` when driving an external coding agent, so the agent's own tools are forwarded and returned as tool calls instead of Unsloth running built-in tools. [source]
- The Codex guide shows `-c 131072` (with `--reasoning on`) as the way to set the context window, and a separate coding-agent command `unsloth run --model ... --disable-tools -p 8888`; with no flags Unsloth picks recommended context length and temperature. [source]
- The Codex guide says Unsloth's self-healing tool calling is on by default and helps when tool calling is unreliable. [source]
- The Claude Code guide's `--disable-tools` command line is `unsloth run --model ... --disable-tools --reasoning off -p 8888` for a coding agent. [source]
- Inferred: `--disable-tools` is the only control the docs document for agent passthrough, so a Mac user binding Studio to loopback for Claude Code should always pass it. [source]
- Inferred: binding to `0.0.0.0` makes passthrough the default, so a LAN-served Studio needs no extra flag for Claude Code but loses the built-in tools. [source]
Children
- No children recorded.