Skip to content

Local AI models through Ollama #20

Closed as not planned
Closed as not planned

Description

@ado11231

What

Run small language models on the Agent's GPU, and use them from the Client as if Ollama were local.

Plan, In Order

  1. Automatic port forwarding #9, automatic port forwarding. Built: run and attach forward Ally's ports, and ally forward 11434 opens Ollama's port on request.
  2. Local models: offer Ollama in the tools step #38, offer Ollama in the tools step, based on the Agent's GPU. Done in feat: offer Ollama and OpenCode in the tools step #131.
  3. Local models: ally models #39, ally models: list models, show what fits in free VRAM, and pull with a warning.
  4. Local models: show the loaded model in the menu bar #40, show the loaded model and its VRAM use in the menu bar.
  5. Local models: offer OpenCode in the tools step #60, OpenCode in the tools step. The code landed in feat: offer Ollama and OpenCode in the tools step #131; only its real test is left.
  6. Local models: a chat window for the Agent's models #57, a chat window for the Agent's models.

Found On October 27, 2026

Rules

  • Ollama runs only on the Agent, and listens only on the Agent's 127.0.0.1. The Client reaches it through the ssh forward from Automatic port forwarding #9.
  • Ally wraps the ollama program. It never runs models itself.
  • Nothing is installed without showing the command and asking once.

Done When

  • On a Client with no GPU, a tool pointed at localhost:11434 uses a model running on the Agent's GPU, and ally models shows what is loaded.

Activity

  1. ado11231 commented on Oct 2, 2026

    @ado11231
    OwnerAuthor

    Update, October 1, 2026: research for Ollama and OpenCode is on #38, and sizing notes are on #39. A chat window is tracked in #57. OpenCode and Ollama together inside slingshot attach need no forwarding, so #38 can start before #9.

  2. added this to the Complete milestone on Oct 16, 2026
  3. ado11231 commented on Nov 2, 2026

    @ado11231
    OwnerAuthor

    Dropped for now. Local models work without Ally commands: run Ollama or OpenCode inside ally attach, or a browser UI through port forwarding (USAGE.md, Use Local Models, in #164). Reopen if a model command is needed later.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions