Home and local model chat

What Home provides

Home combines an in-app local model conversation with Gateway, Remote and Quick Actions panels. You can ask a question without an approved project, or select a project to give the chat relevant context.

The model used by Home runs through a supported local runtime. External ChatGPT and other MCP clients connect separately through Connect; your local model selection does not configure those clients.

Choose a reachable local model

Start Ollama, LM Studio’s local server or llama.cpp’s server on your machine and make a model available in that runtime. Select a reachable model from Home’s model picker.

RepoTunnel discovers models exposed by the three supported local runtimes and refreshes that discovery while the app is running and when it regains focus. Home’s picker lets you choose among reachable models; it does not silently switch to a cloud model when a runtime is unavailable.

• If Home shows No local model, check that the runtime is running and actually exposes a model.
• RepoTunnel discovers models already available through the runtime; it does not automatically install or start that runtime.
• Use the selected model’s real capability information when choosing a task.

Prepare one runtime before using Home

1. Choose one supported runtime and install it from its official site. You do not need all three.
2. In Ollama, download a suitable local model and make sure the local service is running. In LM Studio, download and load a model, open Developer and start the local server. For llama.cpp, follow its server instructions with a compatible local model file.
3. Return to RepoTunnel’s Home screen and select the reachable model. If it is missing, verify the runtime’s server and loaded model before refreshing RepoTunnel.

Ask with the right project context

Select the intended project before asking about its code. Home can include bounded project context, the current file, a selected code fragment you explicitly add, and an existing diagnostic you attach.

Use + Selection or + Error when available. Remove a context chip if it does not belong in the next question. Context may be reduced to fit the selected model, and the chat reports that reduction or a context warning.

• Start with a specific question about the selected project.
• Check the context sources returned with the response.
• Use Shift+Enter for a new line and Enter to send.
• Stop generation if the request is no longer useful.

Opt into edits for one message

Allow edits is a per-message choice for a selected project. It enables that message to propose safe project edits through RepoTunnel’s editing and History path; the toggle resets after sending.

Keep it off while exploring or asking for an explanation. If you enable it, inspect the resulting change or pending request in History and verify the file contents. Workspace permissions and safe-edit validation still apply.

Read the connection panels

Gateway shows whether the local MCP endpoint is online. Remote shows the configured public endpoint or managed client connection and can open Connect.

A running gateway is one layer of the connection. Verify remote work with an authenticated request from the intended AI client against the selected approved project.

Use Quick Actions

Open project registers an existing folder; New project starts the app’s project-creation flow. Create checkpoint saves the selected project’s accessible state, and Safety scan checks its protection status.

Pause AI access temporarily blocks MCP workspace access. Resume only when you are ready for those clients to continue. Phone Access has its own local access and pause controls.

Keep conversations organized

Home stores local conversations and lets you start a new chat, reopen a recent one or delete a conversation from this device. Switching the selected workspace reloads that workspace’s local conversation context.

Deleting a chat and restoring project files are separate operations. Use History when the goal is to recover a file version.