Local agents with Ollama and Aider

OpenVibe is not only for Claude Code. It also runs models that live on your machine, through Ollama, either as a plain chat or behind Aider so they can edit your code. They get their own terminal on the canvas, join groups and take your instructions like any other agent.

But part of OpenVibe relies on mechanisms that only Claude Code provides. This article says what works with a local agent, what does not, and why. It recommends no model and gives no measurements: the right choice depends on your machine and your task.

Two ways to run a local model

Mode	Command run	Access to the repository
Chat	ollama run <model>	None: it is a conversation
Coding agent	aider --model ollama_chat/<model>	Yes: Aider edits the repository’s files

Coding agent needs Aider installed and on your PATH; otherwise its button is greyed out, with “aider not found in PATH”. When Aider is there, it is the mode offered by default.

Either way, the Ollama server has to be running. If it does not answer, OpenVibe says so (“open the Ollama app, then try again”) instead of opening a terminal that would fail.

Installing Ollama and models

If Ollama is not found, Install shows the official command in full, then runs it in a visible terminal on the canvas. The same goes for Aider. You can read what happens, Ctrl+C interrupts it, and nothing is downloaded or run in the background. A project has to be open: the install runs on its canvas.

Then open Launch Ollama… from the function palette (Ctrl+Shift+P) or the agents menu. The list shows the models already installed, which OpenVibe reads with ollama list. To add one, type its name as name:tag and click Download. OpenVibe opens a terminal on the canvas and runs:

ollama pull <model>

You watch the download, you can interrupt it, and a notification says “Model ready” once it shows up in the list. The name is checked before it reaches the shell: only the characters Ollama uses in model names are accepted. The dialog also offers a few suggestions, with their download size; they are starting points, not a ranking.

On the canvas, like the others

An Ollama or Aider agent is a real terminal on the canvas. Everything that does not depend on the agent works the same way:

• Groups. Compose a group picks the agent for each slot: a group can mix Claude Code, Aider and a shell. A team preset remembers the agent of each slot.
• The instruction box. @all, a group or @artemis: your instruction is typed into the terminal. A line starting with / goes to the CLI as is; the suggestions are Claude Code’s, but Aider’s own / commands get through the same way.
• Change review. It reads git, not the agent: you see what Aider changed and send your line remarks back in one message. Aider commits its edits by default, so look at the Since tab (changes since the branch’s base) rather than Uncommitted.
• The rest of the workshop. Code Studio, the git window, notes (plain Markdown files in .openvibe/notes/) and loops ask nothing special of the agent.
• Pausing. For Aider and Ollama, OpenVibe sends Ctrl+C where Claude Code gets Esc.

How OpenVibe reads their state

For Claude Code, the state comes from its hooks: OpenVibe knows exactly when the session is working, waiting for your answer or done. Ollama and Aider have no such hooks. Their state is read from their output: text coming in means WORKING; a second and a half of silence means READY.

That reading has a clear limit: it cannot see WAITING. When Aider asks a question and waits for your answer, its terminal goes quiet, and OpenVibe shows it as READY. What follows from that:

• the “an agent has a question” notice and the Windows notification do not fire for it;
• a loop, which never writes to an agent that waits, cannot know that this one is waiting;
• if a model takes a while to answer without printing anything, its terminal turns READY before the answer comes.

Keep an eye on these terminals, or put them in a group of their own so you can find them at a glance.

What relies on Claude Code

Several features go through what Claude Code offers and other agents do not: its hooks, its status line, its local logs and the MCP tools OpenVibe gives it. They do not work with a local agent.

Feature	Claude Code	Ollama or Aider
Canvas, groups, mentions, review	Yes	Yes
Precise state, including WAITING	Yes, through hooks	No: WORKING or READY, read from output
Missions and orchestrator	Yes	No: a mission’s agents are Claude agents
Model router	Yes	No: it chooses between Claude models
Step timeline	Yes	No
Remote Control	Yes	No
Usage meter, Tokens view	Yes	No: they read Claude Code’s subscription and logs

Nothing stops you from mixing both on one canvas: an orchestrator and its Claude agents on one side, an Aider group on separate tasks on the other, and the same instruction box for all of them. The orchestrator, though, will not hand anything to the Aider group.

The limits, plainly

• No measurements. This article does not compare local models with each other or with Claude. Quality depends on the model, your machine and the task: try it on a real ticket from your repository.
• Your machine does the work. A local model runs on your processor or graphics card, and several local agents in parallel share the same resources.
• The Ollama server has to be running. OpenVibe does not start it for you.
• No missions. To split a large task between local agents, you make the plan, and the instruction box hands it out.
• Early preview, Windows only. OpenVibe is built and tested on Windows 10/11.

What holds for every agent: OpenVibe adds no cloud service and no telemetry, and its only request of its own is the update check, which you can turn off. With Ollama, the model itself runs on your machine.