Claudexor Routes Coding-Agent Quotas

This research library uses AI-assisted source research and drafting. Linked sources support product claims; analysis and proposed exercises are our interpretation. Unless an article documents a test and its results, do not read it as a hands-on review or an independently verified benchmark.

Claudexor can put several coding agents behind one interface, but an adapter alone does not establish that the router knows an account’s remaining quota. That distinction matters when choosing a fallback for an interrupted task.

This is a source review and proposed repository exercise, not a report of a production trial. Updated on September 21, 2026 to clarify quota coverage and the difference between instructions and enforced permissions.

What does the project actually support?

The Claudexor README describes a local-first control plane with adapters for Claude Code, Codex CLI, Cursor CLI and other tools. It separates read-only turns from write work, with patches or manifests available for inspection and cross-harness review.

Its current live-quota section lists Antigravity, Claude and Codex. It explicitly says Cursor has no vendor usage source yet. Do not interpret Cursor adapter support as live Cursor subscription-quota measurement.

The project also distinguishes unknown cost from zero cost. These are project claims and interface descriptions; we have not independently verified every adapter or accounting path.

What should a first trial prove?

Choose a disposable repository with a passing test command and one small defect. Record the starting commit, the failing test and the permitted files. Keep the task and acceptance criteria identical when comparing harnesses.

Ask each candidate to produce a patch. Before applying it to the working branch, inspect the complete diff and run the same check. Record elapsed time through review, including failed attempts and manual correction. A quick first answer can still produce a slower completed change.

Use a short receipt for each attempt:

Starting commit:
Requested change:
Harness and account profile:
Quota observation, source and timestamp:
Files changed:
Test command and result:
Manual corrections:
Measured cost, or unknown:

For a quota-switching trial, record whether the second harness received the same task, working tree and prior findings. A successful switch that loses the failing test is not a successful continuation. Stop if the router cannot explain which account or patch it used.

How do protected paths stay protected?

An AGENTS.md instruction can tell an agent to avoid a directory. It does not remove filesystem access. Reviewers must check the patch, and the chosen runtime must enforce any required write restriction through its permissions or isolation controls.

For the trial, keep credentials and production data outside the checkout. Use an account profile approved for the experiment. If external tools are connected, scope their permissions separately: a repository rule cannot revoke a remote service permission.

Write down an explicit boundary, such as “change the parser and its tests; leave deployment configuration untouched.” Then deliberately request an out-of-scope change in the disposable setup. Verify the actual denial or review rejection. A refusal in one conversation is evidence about that attempt, not a general security guarantee.

Our subagents and skills guide gives the broader repository-instruction context. The local-motion review offers a neighboring local-agent example for comparing where execution happens.

When is a router worth keeping?

Keep it only if the exercise exposes a recurring problem it solves: interrupted work across existing tools, inspectable patch comparison, or a measurable reduction in manual switching. If one harness completes the task reliably, the extra credential and routing configuration has not earned its place.

Does connecting Cursor provide live quota data?

Not according to the current README. Record unavailable quota as unknown and test what the router does with that missing information.

Can a review checklist enforce file access?

No. It can help reject an unacceptable patch. Use runtime controls for access restrictions and verify both layers in the trial.