Delta Claude Agent Setup: Why It Cannot Run on a Cloud Runner

Delta 0.26.0 adds Claude Agent to its collaborative coding workspace, but you cannot run those turns on Delta Cloud Runners. Claude Agent currently runs in the desktop app. If a thread says it “cannot run here,” changing your API key will not solve the execution-location mismatch.

The November 12 release notes also add new-thread creation in Delta Web. Taken together, these changes make the product more flexible, but they do not make every agent available in every environment. That distinction matters when your intended workflow is to start coding from a phone and close your laptop.

Set up Claude Agent on desktop

Open Settings, then LLM Providers, and enable Claude Agent Integration under Agent Connections. Sign in there, then choose Claude Agent under Other Agents in the thread's picker. This is a separate agent integration, not simply selecting a Claude model for Delta's own agent.

The Claude Agent guide describes the integration as beta. An existing Claude Code login can be reused. An Anthropic API key entered for Delta's own provider does not configure this integration, and an exported ANTHROPIC_API_KEY does not silently change a subscription login into API billing.

If you encounter “isn't enabled on this computer,” check the integration setting on that computer. For “cannot run here,” use desktop or switch the thread to Delta's own agent. Make that choice deliberately: preserving a conversation does not mean two agents have identical tools, behavior or execution environments.

This is a documentation-based setup guide, not a report of hands-on testing. Start with a small disposable task before handing either agent a repository with meaningful write access.

Choose the execution environment before the task

The Cloud Runner documentation describes Zed-hosted Linux sandboxes that can keep working while your laptop sleeps. They use the thread's synced worktrees, not every file and credential sitting on your computer. Ignored files, local services and local-only credentials need separate treatment.

Intended job	Practical starting point
Use the Claude Agent integration	Delta desktop with Claude Agent enabled
Run Delta's agent without keeping a laptop awake	Cloud Runner with the project prepared there
Use tools installed on a computer you manage	Personal Runner, with that computer awake and connected

A cloud migration should start with the dependencies of the task, not its prompt. For a frontend change, ask whether the package manager, test fixture and API service exist where the commands will run. A request that succeeds locally can fail remotely without anything being wrong with the model.

Cloud Runners currently exclude Claude Agent turns. Their compute is free until Delta plans roll out, subject to fair use; model charges remain separate. A free runner is therefore not an unlimited free coding agent.

Check which account pays for the model

Delta's plans and usage page separates three routes. Claude subscription access follows the provider's limits. An Anthropic Console login for Claude Agent uses Anthropic API billing. Delta-hosted model usage instead draws from the applicable Zed credits and spending limits.

That makes “I already pay for Delta” an incomplete cost check. Before a long task, identify the selected agent, its signed-in account and the billing route associated with that account. Do not use a second account merely to make a limit message disappear without understanding the resulting charge.

For a team, write these three facts beside a trial task: where it executes, whose model account it uses and how the result will be reviewed. This simple record makes a comparison much more useful than asking which interface feels faster after one successful run.

Browser access does not imply a browser shell

The Delta Web guide distinguishes reviewing synced files from running commands. A browser turn without an execution runner lacks a disk checkout and shell. Choose Cloud for remote command execution, or a connected Personal Runner when the task needs a specific machine.

The same guide still contains older instructions saying top-level threads require desktop, while the newer 0.26.0 release notes announce web thread creation. Treat the dated release notes as the update on that point, and avoid interpreting older desktop-only guidance as a universal limitation.

The useful improvement is being able to keep a shared review workflow while choosing how work gets done. The mistake would be treating a common thread interface as proof that its agents, credentials and computers are interchangeable.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain developer workflows and product changes without hiding the details that decide whether they work.

Book a 30-minute call