[BANADDA]

The first release of the mtensor command, for customers.

• mtensor login stores a key in ~/.mtensor/credentials.json (0600); MICROTENSOR_API_KEY works too.
• mtensor bridge for claude-code, codex, opencode, cursor and vscode: snapshots everything it will touch to ~/.mtensor/bridge//snapshot.json before the first change, writes atomically, registers the bridge with the service when the service supports it, and prints exactly what changed and how to undo it. --model, --key (a key per tool), --dry-run and --force.
• mtensor bridge status, off (removes what Bridge added, keeps other edits made since) and restore (writes every snapshot back byte for byte).
• mtensor models lists general models and certified specialists.

What each adapter writes follows each tool's documented settings: Claude Code ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, a tool header, the model and every tier, and a model picker; Codex a [model_providers.microtensor] table on the Responses API (Codex no longer accepts Chat Completions); OpenCode an openai-compatible provider block; Cursor the OpenAI base URL override and models in state.vscdb; VS Code a customendpoint provider in chatLanguageModels.json. Cursor and VS Code keys go into the editors' own Electron safeStorage format (v10 AES-GCM with the DPAPI-protected Local State key on Windows, AES-CBC with the Keychain password on macOS, the obfuscated fallback on Linux, said plainly); if that cannot be written the editor asks for the key once. Both editors must be closed while their database is written.

Nothing pretends: until a general model is served the bridge refuses with the reason, because the tool would fail on every request; --force writes anyway. Per tool usage shows in status once the service records bridges (gateway routes /v1/bridge/tools come with the next gateway phase). The repository is public under the proprietary licence, like the engine.

🤖 Generated with Claude Code