microtensor-cli

The mtensor command for Microtensor customers. Its first job is Microtensor Bridge: one command points a coding tool you already use at Microtensor models, and one command puts it back.

Install

uv tool install git+https://github.com/microtensor-io/microtensor-cli

or

pipx install git+https://github.com/microtensor-io/microtensor-cli

Python 3.10 or newer. A PyPI release will follow.

Sign in

Create a key at https://www.microtensor.cloud/console/keys, then:

mtensor login

The key is saved in ~/.mtensor/credentials.json, readable only by you. MICROTENSOR_API_KEY in the environment works too.

Bridge a tool

mtensor bridge claude-code
mtensor bridge codex
mtensor bridge opencode
mtensor bridge cursor
mtensor bridge vscode

Options for every tool:

Option	What it does
--model <id>	The model to use. Default mt/auto, which picks a model for each request. mtensor models lists the choices.
--key <mt_sk_...>	Use a different key, for example one key per tool so usage is split by key.
--dry-run	Print what would change and write nothing.
--force	Write even when the tool is running, or when no general model is served yet.

Before the first change, Bridge saves a snapshot of everything it is about to touch in ~/.mtensor/bridge/<tool>/snapshot.json.

What each tool gets

Tool	File it edits	What changes
Claude Code	~/.claude/settings.json	env.ANTHROPIC_BASE_URL, env.ANTHROPIC_AUTH_TOKEN, a tool header, the model and every model tier, and a model picker of Microtensor models. Start a new session afterwards.
Codex	~/.codex/config.toml	A [model_providers.microtensor] table using the Responses API, model_provider and model. Restart Codex afterwards.
OpenCode	~/.config/opencode/opencode.json	A provider.microtensor block with the models and model. Restart OpenCode afterwards.
Cursor	Cursor's state.vscdb	The OpenAI base URL override, your key in Cursor's own encrypted store, and the Microtensor models. Quit Cursor first.
VS Code	chatLanguageModels.json and VS Code's state.vscdb	A custom endpoint provider with the Microtensor models, and your key in VS Code's secret storage. Quit VS Code first.

Limits worth knowing:

• Claude Code sends every request to Microtensor while bridged, background tasks included, so every tier points at the chosen model.
• Cursor routes custom key requests through its own servers, uses them for chat only, and some agent features refuse custom keys. Tab completion stays on Cursor.
• VS Code uses Microtensor in Chat; inline suggestions stay on Copilot.
• On Linux, Cursor and VS Code keys are stored the way those editors store them without a keyring: obfuscated, not encrypted.

Check and undo

mtensor bridge status
mtensor bridge off claude-code
mtensor bridge restore

• status shows each tool, whether it is bridged and to which model, and its requests once the service records them.
• off <tool> removes what Bridge added and puts back the values you had before, keeping any other change you made since.
• restore writes every saved snapshot back exactly as it was before the first bridge.

Models

mtensor models

Lists what Microtensor serves: general models, which coding tools need, and certified specialists, which answer one task each. Until a general model is served, mtensor bridge refuses and says so, because the tool would get an error on every request; --force writes the settings anyway.

Development

pip install -e ".[dev]"
ruff check . && mypy microtensor_cli && pytest -q

Docs: https://www.microtensor.cloud/docs/api/getting-started/bridge