AI Chat

A turn, step by step

1. You send a message. Attachments and images can be included.
2. The agent reports each tool call as it happens: reading a file, searching the project, running a command.
3. An action that changes something waits for your decision.
4. When the turn ends, the conversation is saved locally.

For a requested action, Always stops asking for that action in the current chat, Once allows it a single time and Deny skips it. Autopilot answers these prompts automatically for the whole turn.

Agents on desktop

Each row is a local agent integration that Genexy runs on the desktop device through the agent’s own CLI or adapter, so none of them run on Android or iOS. Where the agent offers its own login, Genexy uses that sign-in session and stores no key. Claude Code runs through its ACP adapter and authenticates with a key only.

OpenCode and Cline are open-source agents that route to many model vendors. In the settings you connect a provider such as Anthropic, OpenAI, Google Gemini, OpenRouter, DeepSeek or xAI with its key; Genexy writes it to the agent’s own credential file, so the provider also works when you run the CLI in a terminal. Anything not in the list, such as a subscription login or a local server, is connected with the agent’s own login (opencode auth login, cline auth), which the settings open in a terminal.

Several agents can be used side by side in one project, each in its own chat.

Genexy does not provide or resell AI access, and none of it is behind PRO: every agent and provider here is part of Free. You bring the account or the key, and requests go straight from the IDE to that provider.

Self-hosted and OpenAI-compatible models

Currently available free providers are preconfigured so AI chat works immediately, on every platform, without an API key. They are third-party services, so their availability, rate limits and model lists change outside the Genexy release cycle.

You can also add your own OpenAI-compatible endpoint with a name, base URL and optional API key: a model you run on your own hardware with Ollama, LM Studio, vLLM or llama.cpp, a service such as OpenRouter, or a company gateway. Any model that speaks the OpenAI-style chat interface can be used, on every platform. Tool support depends on the endpoint and the model: if a model refuses tools, Genexy continues the turn as text-only chat and tells you that project actions are unavailable.

Switching mid-conversation

The selectors under the composer change the agent, provider or model without starting over. The existing context is kept, the switch is recorded in the conversation, and Genexy warns you when the new target cannot use tools.

The composer also shows how much of the model’s context window the conversation currently occupies, so you can start a new chat before quality degrades.

Genexy MCP

Every agent gets the basic tools for free, in the IDE and through the OpenAI-compatible providers alike: reading, searching and editing files (read_file, search_code, str_replace, apply_patch…), running commands in Genexy Shell, checking syntax and reading Git. That is what makes AI work on a phone too.

With PRO, Genexy also gives the agent its own MCP tools, so it can use most of the IDE: run configurations, the debugger, bookmarks and TODO, terminals, language servers, Docker, connections, libraries, the code generators and more. Tools of a feature with its own plan, such as Docker, need that feature as well. The tools run inside the IDE on your device; they are not a public or remote endpoint.

History and data flow

Conversations and attachments are stored on the device. On desktop they live under:

~/.genexy/chats

Mobile operating systems map the same application storage into their sandbox.

Requests go from the IDE straight to the selected agent or provider. Genexy does not proxy AI traffic through its own servers, which also means the provider’s terms, retention and training policy are what apply to your prompts.