Use desktop models on a laptop

Can my laptop use a model on a more powerful PC?

Yes. Lyre can use a selected, eligible model from another connected computer while an agent works in your laptop's project. The desktop is the lender: it runs model inference using its configured model source. The laptop is the borrower: it owns the agent, workspace, project files and supported project tools.

This can be useful when a local model is set up on your desktop and you want to edit a laptop checkout without running that model on the laptop. The laptop still needs the dependencies and resources for its own tools and app. There is no promised speedup: model, network and project-tool performance all affect the result.

Choose where the work should happen

Borrowing a model does not move your build to the desktop or synchronize repositories. A dependency missing on the laptop remains a laptop setup task. If the whole project needs to run on the desktop, choose that computer's workspace instead.

Prepare two connected computers

A Developer connection may discover and use only the lender's permitted models; it does not need a project on the lender. App-only access does not permit model borrowing or agent work. Known native Codex and Claude account sources are excluded from Developer lending. Supported Owner account routes are separate and keep their current source requirements; this is not a way to transfer every provider subscription.

Let the lender choose what to share

1. Open model sharing on the desktop

   As the lender's owner, open Settings›AI & agents›Orchestration›Remote agent here, then Models across computers. Under Share this computer's models, follow its permissions link and choose who may use the models. Keep the computer name in view. Sharing setup.

2. Set Developer grants and limits when needed

   On a host offering Developer lending, choose eligible entries in Models lent to developers. Prefer individual models when a whole provider is unnecessary: an all-models provider entry can include models enabled later. Review Limits and any per-person overrides, then Save lending settings.

3. Check the authority you are testing

   Owner use bypasses Developer lending grants and limits. A test through your Owner connection does not establish what a Developer can borrow. Use the intended Developer connection to verify that only the granted models are available. The grant list applies to eligible Developer use; limits can be set per person.

Limits bound concurrent requests and requests per lender-local day; they are not a promise of inference capacity or a spending cap at an external provider. Provider charges and account terms still apply. Keep the model endpoint and provider credentials on the desktop rather than copying a private key into the laptop project or exposing the model server publicly.

Select the model and work in the laptop project

1. Save the laptop's source-model choice

   Select the laptop in Settings and open Settings›AI & agents›Orchestration›Remote agent here. Under Use models from other computers, find the desktop's card, choose Choose models, then the provider and available model. Wait for the selection to save. A saved offline choice is not current readiness.

2. Create the agent in the laptop workspace

   Open the laptop project and Orchestrate›Remote agent here. Confirm Runs on names the laptop and the intended project, while Model from names the desktop. Select the saved model. Read the reported tool capabilities before Start remote agent.

3. Start with a read-only project task

   Use the request below, then compare the response with the actual README on the laptop. If it is correct, ask for one specific, reviewed typo correction in that file. Inspect the diff on the laptop and run the relevant existing checks there. Treat the request as an example, not a recorded model result.

Read README.md in this project and summarize how to start it.
Do not edit files or run commands. If README.md is missing, say so.

Understand what crosses the connection

The lender's saved model endpoint and provider credentials stay on that computer. Inference requests and replies cross the authenticated bridge. Instructions, relevant file contents and tool results used by the model can therefore reach the desktop and any external provider it uses. Local ownership of the project does not mean zero data transfer. Review security and provider disclosures before using sensitive work.

The managed Remote agent here route uses the laptop's permitted project filesystem and, when the host advertises support, its agent-owned terminal sessions. External MCP tools are not supplied by this managed route. Read the capability line; an older execution host may offer file tools without terminal commands.

A separate Owner option named after the source computer with OpenCode runs the laptop's installed OpenCode with its native tools, approvals and MCP configuration. It needs laptop Owner authority and is not the same managed tool boundary. Follow remote-model tool choices and tool setup before choosing it.

Recover the source or an interrupted request

The desktop is offline or its models cannot be chosen

Restore both host connections, check the desktop card and its provider readiness, and verify the intended model is still granted. Older hosts can require an update. Saved selections do not override source availability or permission.

A request hits a limit or the provider refuses it

Read the actual lender or provider error. The owner can inspect Limits and In use now; provider quota and account restrictions are separate. Reconnecting does not replenish a quota or make a restricted account eligible.

The model asks for a tool the route does not support

Read the laptop agent's capability line and approval state. Use its supported tools or deliberately choose an eligible executor; borrowing a model does not grant arbitrary terminals, MCP or desktop control.

A request disconnects or its result is unclear

Check the laptop conversation, file diff and owned terminal state before retrying. Source loss or revocation ends affected inference; Lyre does not silently choose another source and replay tools. Existing file edits are not automatically rolled back. Source recovery.

Stop a request or revoke lending

On the lender, In use now›Stop and its Stop requests confirmation cancel that person's active requests. Their access stays on. History records request metadata and outcomes; it is not a transcript of the borrowed project's prompts or files.

To withdraw a model from Developer lending, remove its relevant entry under Models lent to developers and choose Save lending settings. This changes the shared Developer grant list; Owner use remains separate. To withdraw the device connection itself, the owner revokes that connection through Pair device. Removing a saved choice on the laptop only stops selecting it there; it does not revoke the lender's grant.

Verify withdrawal through the affected Developer connection: an in-flight request must stop delivering, and a fresh request for the withdrawn model must be denied. Inspect the laptop's current work before restarting elsewhere. Revocation cannot erase context already sent to a model or undo completed tool effects.