Cline released a beta desktop application for Mac and Windows that runs coding agents against open-weight models rather than a single proprietary subscription service. The company, whose VS Code extension has been installed by more than 11 million developers, says the standalone app extends its existing open agent harness to people who do not work primarily in a terminal, adding parallel task sessions, scheduled automations and a plugin marketplace.

The app’s most notable feature lets a developer import an in-progress session from Claude Code or OpenAI’s Codex and keep working on it with a different model. Cline frames this explicitly as a fallback for when a subscription’s usage quota runs out. That framing positions Desktop less as a head-on competitor to Anthropic’s or OpenAI’s coding agents and more as a release valve for teams already paying for one of them.

Developers can select from more than 300 models through Cline’s own provider, tap open-weight models through a service the company calls ClinePass, run models locally, or connect an existing API key to over 50 providers. Cline’s two-stage workflow, Plan and then Act, can run each stage on a separate model, a level of control that subscription-bundled coding agents typically do not expose to users.

Cline says it rebuilt the underlying agent harness around a new SDK, rewriting its prompts, simplifying the agent loop, and changing how models locate and invoke tools. That rebuild, not the desktop interface itself, is doing the real work: which model a developer picks only matters if the harness around it manages context and recovers from tool errors consistently.

The company published Terminal-Bench 2.0 pass@1 scores comparing its harness against two other open-source coding-agent harnesses, Hermes and OpenCode, across four open-weight models. Kimi K3 posted the highest result under Cline’s harness, at 82.02 percent, versus 71.9 percent under Hermes and 76.4 percent under OpenCode on the same model. Cline says these figures were reproduced from its own September 2 harness-upgrade post rather than generated as a fresh, Desktop-specific evaluation, and no independent benchmark accompanies the release. Vendor-run comparisons of a company’s own harness against rival open-source tools are the vendor’s claim, not a neutral test.

Desktop also ships three smaller additions: a Schedule tool for recurring cron-job automations, such as checking a repository every night or refreshing documentation once a week; web search for pulling outside context into a task; and voice input for dictating instructions instead of typing them. All three extend the existing harness rather than introduce a capability unique to open-weight models.

For teams already paying for Claude Code or Codex, Cline Desktop is worth testing specifically as a cost hedge: the ability to move a quota-capped, stalled session onto a cheaper open-weight model without losing prior context is the one feature a closed-source rival cannot easily replicate. The Terminal-Bench comparisons against Hermes and OpenCode are self-reported and specific to Cline’s own harness, so operators evaluating open-weight coding agents should benchmark their own workloads before shifting budget away from existing subscriptions.

Reported by Cline on its company blog, September 14, 2026.