Coding agents do their best work on a machine that has the code, the build tools and the credentials. For most teams that machine is a server, not a laptop. Olivares AI runs Claude Code, Codex and Grok Build on that server and gives you one web console and one CLI to start, follow and stop their sessions.
Install the engine
Olivares AI is a single program with the console built in. On Linux or macOS one command installs it, and it verifies the release before it installs anything:
curl -fsSL https://olivares.ai/olivares/install.sh | sh
olivares quickstartquickstart prints the console address and a one-time token to create the first administrator. Docker images, Debian, RPM and Alpine packages and a Homebrew tap are also available; the install guide lists every method.
Connect a tool with the account you already have
Open AI tools in the console. It installs Claude Code, Codex and Grok Build and runs each tool’s own login flow, so you sign in with the subscription your team already pays for. Every account you sign in becomes a separate tool instance, and the console shows the plan, usage windows and models that the tool reports.
If you prefer API keys, add them under Providers. A local model works too: Ollama needs only a reachable endpoint and a model name.
Start a session
Choose Sessions → New session and pick a ready tool. The engine fills in the profile and the project folder. Enter the first message and press Start.
The same works from a terminal. session start takes the folder and the first prompt:
olivares login
olivares session start ~/work/api "Explain this project." --name api-review --detach--detach sends the prompt and returns without waiting for the reply. The session keeps running on the server, even after you close the terminal. Follow its output again from any terminal with olivares session follow api-review. With Business, the console’s live terminals and olivares session send and resume let you keep steering it.
For a git repository, Work in a new git worktree gives the session its own branch and working directory, so two sessions never edit the same checkout. Cleanup refuses to discard uncommitted or unmerged work unless you say so explicitly.
Approve, interrupt, stop and resume
Turn on Ask before each action and the tool waits for you before it acts. The request appears in Approvals, where you approve or deny it.
Interrupt ends the current turn and keeps the conversation. Stop ends the process. With Business, Resume restarts the same session on the same profile, account and folder, and checks credentials and policy again first. A session never switches to a different login on its own: if its account is no longer available, the launch is refused.
Rules that apply to every session
Because every session starts from the engine, the same rules apply to all of them. Olivares AI checks actions at four deny-closed enforcement points: inside Claude Code, in the model proxy, at every MCP tool call and between agents. With Business, budgets per team, agent or model warn, slow down or stop spending. Each decision goes into a signed, tamper-evident log.
Claude Code and Codex sessions bound to a provider key reach only their provider’s endpoint, through a per-session proxy, or they do not start. Sessions that sign in with an account, and Grok Build and OpenCode sessions, are not isolated this way: they can reach any host your network allows.
Try it
Community is free and open source, with unlimited users. Install it from the GitHub repository, or read how it works first. Business adds what larger and regulated organizations need, with a commercial license.