Juggler Turns Coding Agents Into a Workbench

Juggler, by Julian Storer, is a visual workbench for coding agents. Its current README describes persistent conversation trees and inspectable model transactions, including prompts, messages, tool definitions and outputs. It can connect to providers including Claude Code and Codex. These are recorded interactions, not a view into hidden model reasoning.

Updated 21 September 2026: consolidated the three Juggler articles, checked the current README, and corrected remote-access guidance. The procedure below is a proposed evaluation, not a hands-on result.

Check where the session runs

The project distributes a desktop client and a headless server. The documented server listens on localhost by default. Pressing p or starting with --public enables LAN access without a password; the README explicitly restricts that mode to trusted networks. Official binaries also have separate WAN components. Check the exact build and access mode before exposing a session. Source: Juggler README.

For a first trial, keep the server local and use a repository without production credentials. A Git branch makes a patch easier to discard, but it does not isolate the process from credentials or other files available to its account. Choose the execution environment and credentials separately from the branch.

Give the reviewer a concrete question

Pick an existing parser bug with a small failing example. Write down the expected behavior, permitted files, dependency policy and stopping condition before starting. Ask the agent to repair the behavior and run the relevant test. Avoid a broad request to improve the entire repository: it makes the review comparison hard to interpret.

After the run, ask a second person to reconstruct the change from the session and repository. They should identify the input that motivated the edit, the tool action that changed each file, any abandoned approach, and the command that checked the final result. Compare their account with the actual diff, including untracked files.

A readable session is useful only if its evidence agrees with the repository. A recorded passing command can still refer to an earlier patch. Check the order of the last edit and the last test. If the agent changed the test to accept the bug, the green result does not satisfy the original task.

Keep a portable handoff

Record these fields in the pull request or local review note:

The same handoff works when the final review happens in Cursor, Claude Code or a normal Git diff. Repository instructions can request it; those instructions are not an access-control mechanism. Restrict write-capable integrations through their actual credentials and permissions.

Keep Juggler if this exercise makes a difficult session easier to review. If the ordinary diff and command output already answer every question quickly, there may be little benefit in adding another interface. The decision should come from the review effort and missing evidence, not the appearance of the session tree.