CLI connections reuse an already installed and authenticated command-line tool. Opening a CLI configuration automatically fills an empty executable field when Vopit finds an installation. You can also select an executable manually; automatic detection preserves a path you entered. Save settings to keep the detected path. Use Verify installation and Test response before assigning the connection to a production task.
| CLI | Terminal sign-in shown in the app | Windows executable |
|---|---|---|
| Codex CLI | codex login | codex.exe or the standard npm codex.cmd shim |
| Claude Code | claude auth login | claude.exe or the standard npm claude.cmd shim |
| Mistral Vibe | vibe (configures its credentials) | The native vibe-acp executable; .cmd and .bat shims are rejected |
| Mammouth Code | mammouth (complete its setup) | The native mammouth executable; .cmd and .bat shims are rejected |
| Antigravity CLI | agy (sign in once) | The native agy.exe, detected on PATH or in %LOCALAPPDATA%\agy\bin |
Codex CLI
- Sign in to Codex CLI in a terminal according to your organization’s approved method, then select Codex CLI in Settings → Connections → Accounts & CLI.
- Verify the installation, refresh the model catalogue, and choose a model and supported reasoning effort.
- Grant an agent native Project inspection before enabling Editing & validation or Git changes. Web research is independent. These selections are capabilities, not proof that a given task is safe.
- Vopit opens native app-server sessions for stages. Codex owns its internal turns and compaction; Vopit’s maxTurns does not limit those internal turns. When the agent may edit, the project folder and its linked folders are the writable roots of the Codex sandbox.
Claude Code CLI
Configure the local Claude Code executable, sign in through the CLI, then verify and test a response. Vopit does not provide a direct claude.ai account connection; use Claude API for API billing. Choose a Connection default model from the catalogue, where aliases such as Opus follow the latest model of their family and pinned IDs stay on one version, or keep Use CLI configuration. Set a Default reasoning effort (low, medium, high, xhigh or max); an agent can override both the model and the effort. When the Ultra-fast or Balanced speed mode sets an effort, it takes precedence over the agent and connection values.
Claude Code uses its native file, search, editing and command tools in the selected project. Linked folders are added to the session when the agent has project access, and Vopit still checks every path. Vopit requests stage access and handles native permission prompts, records tool activity and file changes, validates the final result, and retains the native session for Resume. Vopit runs Claude Code in restricted mode with all hooks disabled and a strict MCP configuration: only Vopit’s own bridge and the MCP servers explicitly granted to the agent are loaded. Use a current official Claude Code release.
Grant Claude Code MCP servers to an agent
- Add the server to Claude Code as usual, for example with claude mcp add. Vopit lists servers declared in your user configuration (~/.claude.json, or .claude.json in the folder set by CLAUDE_CONFIG_DIR), in the project’s .mcp.json, and in folder-local declarations.
- Edit an agent whose connection is Claude Code. In its capabilities, MCP servers shows each server with its source: Your Claude Code configuration, This project's .mcp.json, or Declared for one folder. A folder-local declaration can be granted in any project.
- Switch on only the servers this agent needs, then Save agent. Each grant is stored as mcp:<server> in the agent’s tools and counts toward the 20 selectable actions per agent. The name vopit is reserved for Vopit’s own bridge.
- During a stage, each call to a granted server’s tool asks for Vopit NETWORK permission. Other servers stay unavailable. A server removed from your configuration is shown as No longer in the Claude Code configuration until you switch it off.
Mistral Vibe CLI
- Install the official Mistral Vibe CLI and run vibe in a terminal to configure its credentials. Vopit reuses this CLI configuration; importing a Mistral API key into Vopit does not sign in to Vibe.
- In Settings → Connections, select Mistral Vibe. Use automatic detection or select the vibe-acp executable installed alongside vibe. On Windows, select its native .exe; a .cmd or .bat shim is rejected.
- Choose Verify installation, then Test response. Leave the model blank to use the Vibe configuration, or enter a model alias configured in Vibe. A test response uses the connected account.
- Set Mistral Vibe as the workspace default or as one agent’s override. Vibe uses its native tools and session over ACP, with granted Vopit app actions as a complement. Vopit forwards native permission requests; Vibe also applies its own installed configuration and mode.
Mistral API
For an API connection instead, choose Mistral API, import a key file, choose Save & load models and select an available model. The default endpoint is https://api.mistral.ai/v1. This connection uses the Vopit tool loop, not the Vibe CLI. API credentials stay in encrypted OS storage.
Mammouth Code CLI
- Install Mammouth Code from its official site, then launch mammouth in a terminal and complete its setup. Credentials and billing follow the connection configured in the CLI. A Mammouth API key imported into Vopit does not configure the CLI.
- Open Settings → Connections → Accounts & CLI → Mammouth Code. Automatic detection searches PATH and your home .mammouth/bin directory. On Windows, select mammouth.exe if it is not detected; a .cmd or .bat shim is rejected. Save settings to keep the path.
- Choose Verify installation to check the executable and list its provider/model IDs. Leave Connection default model blank to inherit the CLI model, or select an ID from the catalog. Test response sends a short prompt using the CLI credentials and can consume credits.
- Set Mammouth Code as the workspace default or choose it for one agent. Vopit starts mammouth acp, supplies an isolated read-only or editing profile for the stage, forwards native permission requests, records tool activity and modified files, and validates the final handoff. Pause stops the process; Resume reopens the retained session when the provider, project, model and capabilities still match.
Antigravity CLI
- Install the Antigravity CLI, then run agy in a terminal once and sign in with a Google account, a Gemini Enterprise account or a Gemini API key. Vopit reuses this sign-in and never copies its credentials.
- Open Settings → Connections → Accounts & CLI → Antigravity CLI. Automatic detection searches PATH and, on Windows, %LOCALAPPDATA%\agy\bin. Save settings to keep the path.
- Choose Verify installation to check the executable and list the models from agy models. Each thinking level is its own model, for example gemini-3.8-flash-high. Leave Connection default model blank to keep the CLI model. Test response sends a short prompt and counts toward your Antigravity usage.
- Set Antigravity CLI as the workspace default or choose it for one agent. Vopit starts agy acp, uses its default mode for stages that change the project and its plan mode for read-only stages, and applies the speed mode, agent or connection effort when Antigravity offers that level (low, medium or high). Every edit or command Antigravity asks for goes through Vopit permissions. Pause stops the process; Resume reopens the retained session when the provider, project, model and capabilities still match.
- Antigravity generates images through its image-generator subagent, which runs in the background and reports only when the next prompt arrives. When a stage starts a subagent, Vopit keeps the session open and checks back (after 10, 20 and 30 seconds, then every minute, for up to 10 minutes) until the result arrives; the agent activity says so. A subagent cannot ask Vopit for approval and stalls on a command, so Vopit tells Antigravity to use subagents only for capabilities it lacks, without commands, and never in read-only stages.
The Mammouth profiles are process-local: Vopit does not rewrite your personal CLI configuration or copy its credentials. Native subagent delegation and access outside the project are disabled in these profiles. Granted Vopit actions remain available through the MCP bridge, including Cranks and infrastructure. These controls are not an OS sandbox; an authorized shell command still runs with the current user’s permissions.
Mammouth API
For direct model calls, select Settings → Connections → API connections → Mammouth AI. Get API key opens Mammouth API settings. Import a .txt or .key file containing only the key, then choose Save & load models. Select a text model with tool calling support. The endpoint is https://api.mammouth.ai/v1. This connection uses Vopit tools and Mammouth API credits, independently of the CLI configuration. Optional cost estimates must match the selected model.
Project inspection and shared actions are different
In native project access, the switch labeled Git changes enables inspection of the working-tree diff. Branch creation, request commits, synchronization and history changes are separate Git actions with separate gateway permissions. Infrastructure access also requires explicit target grants. Vopit serves Git, infrastructure, board and Crank actions, search_code_graph and the project memory tools to every native CLI through its app bridge, filtered to the agent's grants. Native stage access does not silently approve those shared actions.
Claude Code runs restricted: it loads only the Vopit bridge and the MCP servers granted to the agent (--strict-mcp-config), its hooks are disabled, and its permission prompts go through Vopit approval. For an agent that runs on Claude Code, the agent editor lists under MCP servers the servers declared in your Claude Code configuration, the project's .mcp.json or a folder's local configuration. Switching one on grants mcp:<server> to that agent; each call of its tools asks for NETWORK permission. Other servers stay unavailable.