Unknown cost is never reported as zero
Multi-harness control plane for Claude Code, Codex, Cursor, and OpenCode: quota-aware rotation across multiple Claude/Codex subscriptions, shared thread context, and cross-model review.
At a glance
- What is it?
- A local-first control plane that puts Codex, Claude Code, Cursor and OpenCode behind one typed interface, where read-only turns resume the vendor's own session and write turns land as inspectable patches. The subscription-quota layer is the interesting part, and the release verification chain is the surprising one.
- Who is it for?
- Adopt claudexor if you pay for more than one coding-agent subscription and want one place that knows which one has quota left, since the credential profiles and typed quota facts are the layer no vendor ships. Do not adopt it expecting one agent with a friendlier interface, because most of its value is in the audit and arbitration machinery around the vendors rather than the chat.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- razzant/claudexor525 ★ · +81
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| razzant/claudexor | 444 | 525 |
81 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
A read turn resumes the vendor session, a write turn produces a patch
The interface is described as a chat of turns, and the turn type decides what happens.
Read-only questions resume the vendor's own native session. Write turns land as inspectable patches or complete file manifests.
That split is the core design choice. Resuming the native session means a read question does not re-establish context from scratch, so the vendor's own compaction and caching behaviour still applies. Making writes produce a patch or a manifest means the output is an artefact rather than a description of one, and a diff can be reviewed before anything touches the working tree.
The third mode is races, where harnesses are pitted against each other with cross-family review. Two different vendors solving the same task and a third judging the difference is the mechanism the README calls best-of-N with independent reviewers and arbitration.
What ties it together is the claim about facts. Cost, quota, web evidence and auth route are described as typed facts you can audit, never a vibe, and the project's own version of that principle is a single rule: unknown cost is never reported as zero.
That rule matters more than it sounds. A tool that cannot price an operation should say so, and a layer that sits between you and several billing systems is exactly where a confident-looking zero would appear. Everything runs on your machine, files are the source of truth, and there is no telemetry.
Credential profiles bind named subscriptions, and Cursor has no quota source
Since version 2.1 the project has had credential profiles, which it describes as named Antigravity, Claude, Codex and Cursor subscription bindings side by side, each with Claudexor-scoped state and platform-declared credential custody.
Four words in that do real work. Named means a profile is a label you chose rather than an account it discovered. Side by side means several live at once. Claudexor-scoped state means the tool keeps its own bookkeeping per profile. And platform-declared credential custody is the one that says plainly where the credential lives, by whatever rules the platform sets.
On top of that there is live subscription-quota tracking, plus an opt-in policy that rotates a spent account out of the way when a vendor limit is reported. The README is precise about coverage: the tracking applies to the harnesses that have a vendor usage source, which it names as Antigravity, Claude and Codex, and it says Cursor has none yet.
That single sentence is the most useful thing in the section. Rotation cannot work for an account whose usage the vendor does not expose, so the feature's reach is exactly the set of vendors that publish a usage signal. Anyone planning a rotation policy across several tools should check which of them are in that set before relying on it.
The opt-in framing matters as well. Rotation is not automatic; it is a policy you turn on, which means the default behaviour when an account runs out is to stop rather than to switch.
Installing a vendor CLI requires proving where the binary landed
The most detailed passage in the README covers managed installation of vendor CLIs, and it reads like a supply-chain contract rather than a feature.
Host integrations may install one vendor CLI into a managed local toolchain:
claudexor harness install <harness> --target local --yes --jsonExact npm pins install under `~/.claudexor/node`. Cursor remains unpinned, so the JSON receipt records the downloaded installer's SHA-256 and byte length. And every successful executed receipt records the absolute installed launcher together with its verified version, with the rule that a zero-exit installer without that postcondition is a typed failure.
That last clause is the whole point. An installer that exits zero tells you the process finished, not that the thing you asked for exists at a known path. Requiring a recorded absolute path and a verified version turns a silent no-op into a typed error.
There is a specific Windows detail that shows how carefully this was thought through. The launcher there is the vendor's own `codex.exe` inside the pinned npm platform package, never npm's `.cmd` shim, and the same managed-root path is what doctor, login and runs resolve. Codex is named as the one vendor with that verified image in this release, and the others refuse typed.
So on Windows, one vendor is fully verifiable today and the rest decline rather than pretend. A tool that reports which of its integrations it can actually verify is more useful than one that reports all of them working.
The release chain fails if a committed schema is stale
The root package file contains a verification chain long enough to be worth reading in full, because it shows what this project considers a release blocker.
It builds, typechecks, typechecks the tests separately, runs the tests, and regenerates the JSON schema. Then it checks that `git diff --exit-code` passes on the generated schema directory, which fails the release if regenerating produced a change. It then asserts that `git ls-files --others --exclude-standard` returns nothing for the same directory, so an untracked generated file also fails. After that it validates the generated schemas, runs the docs check, a concept gate scoped with `--since-last-tag`, and the staged, retired and sensitive-resource checks.
The `git diff --exit-code` step is the one worth stealing. Generated schemas drift silently in most projects: the code changes, the generator is not re-run, and the committed artefact quietly describes the old interface. Turning that into a release failure is cheap.
The other checks are single-purpose scripts with narrow names: a docs truth check, a staged field check, a retired key check, a sensitive resource ownership check, a release workflow check, a concept gate, and a real harness battery that runs against the actual vendor tools rather than mocks.
Two other scripts show the shape of the testing. There is a separate `canary` script running Vitest against `vitest.canary.config.ts`, a second config distinct from the main test run. And `knip` is wired in to find unused code. Formatting is Prettier, builds and typechecks run through Turbo, versions are cut with changesets, and the package pins pnpm at 10.34.5 with Node 20.19 or newer.
You log in through Claudexor, not through the vendor CLI
The prerequisites section contains one requirement that changes who holds your credentials.
You need at least one logged-in vendor CLI: `codex`, `claude`, `cursor-agent` or Cursor's `agent`, `opencode`, or `agy` for Antigravity with a Gemini subscription. Or a provider API key, since adapters accept `OPENAI_API_KEY`, `ANTHROPIC_API_KEY` and similar as fallbacks, and the raw-API route needs only a key.
The instruction that follows is the significant one: log in through Claudexor, not the bare vendor CLI.
That is what makes the credential profile model coherent. If the login happened in the vendor's own tool, the token would be in that tool's storage and Claudexor would be reading someone else's state. Routing the login through the control plane is how each profile gets Claudexor-scoped state and a declared custody location.
The rest of the prerequisites are ordinary but specific. Node 20.19 or newer, because the daemon, the CLI and every surface run on it. pnpm through corepack. Git, needed for Git-backed worktrees, candidate envelopes and source-control delivery, although explicit directory execution supports direct work and selected-input copies without it. A Doctor command reports Git availability, and the app's workspace Git check explains whether a Git-backed shape is admitted.
Platform support is split cleanly. macOS is required for the desktop app, while the CLI and daemon also run on Linux. So a Linux user gets the control plane without the desktop surface.
Two binaries from npm, and a DMG that bundles its own daemon
The install section offers three routes with different amounts of ceremony.
The npm route installs both binaries at once:
npm install -g claudexor
claudexor doctorBoth the `claudexor` and `claudexord` binaries are installed, so the CLI and the daemon arrive together, and `doctor` is the diagnostic to run immediately afterwards rather than after something breaks.
On a Mac there is a signed and notarized DMG, described as installing like any ordinary Mac app with no Gatekeeper warnings. You download `Claudexor-<version>.dmg` from Releases, drag `Claudexor.app` into Applications, and open it. The app starts its bundled Claudexor engine, and onboarding then checks the external Git and harness route needed by the work you selected.
Bundling matters here. The app ships its own daemon runtime and starts it on launch, so installing the CLI is only needed for terminal use. Two daemon installations on one machine is the kind of thing that produces two state directories, and the design avoids it by making the app self-sufficient.
There is also an honest note about an old build. The v1.0.0 DMG was unsigned, and the README tells anyone who kept it to either upgrade or approve it through System Settings, Privacy and Security, Open Anyway. That is a small paragraph that removes a real support question.
The remote SSH path, where the macOS app runs a thread on a Linux or macOS host, is named in the table of contents but the text describing it is cut off before it starts.
Retired verbs hard-error instead of aliasing
The README states a compatibility policy that is more specific than most projects bother with.
Retired verbs and mode ids hard-error with the new spelling instead of silently aliasing.
Silently aliasing an old name to a new one is the friendlier choice for a user and the worse one for a codebase, because the old spelling keeps working in scripts long after anyone remembers what it does. Hard-erroring turns a deprecation into a compile-time-style failure at the boundary.
The README points at a section titled Stability at 2.0 for what counts as a stable contract against what remains experimental, and the current status is stated as v3.17.2.
The project also runs inside another one. It is described as the exact-pinned delegated execution layer inside Ouroboros, a persistent self-developing agent. The division of responsibility is clean: Ouroboros owns tasks, memory, review and final integration, while Claudexor runs the connected coding harnesses and returns durable execution evidence. An exact pin means the two do not drift.
There is one more capability with explicit non-guarantees attached. Embedding callers can make a single model request through a managed Codex subscription using the same account as the agents, supplying their own system prompt and tools and executing those tools themselves. The README is careful to say this is a typed engine operation and not a public OpenAI-compatible server or a second agent loop, that model catalogs and context limits are account-specific, and that subscription access does not guarantee zero incremental charges or enable provider-paid credits.
Finally, the README credits Praxis Relay and CLIProxyAPI for prior work on subscription-backed model transports, and states that they informed the design while neither runs as an embedded relay or owns credentials in this integration.
Editorial conclusion
Adopt claudexor if you pay for more than one coding-agent subscription and want one place that knows which one has quota left, since the credential profiles and typed quota facts are the layer no vendor ships. Do not adopt it expecting one agent with a friendlier interface, because most of its value is in the audit and arbitration machinery around the vendors rather than the chat. Two things to check before you connect anything: that the vendor CLIs you want are among the ones it can verify a local install for, since only Codex currently has a verified image on Windows, and whether you are comfortable logging in through claudexor rather than the vendor's own CLI, since that is where credential custody is decided.
Frequently asked questions
What is Claudexor?
A local-first control plane that runs Codex CLI, Claude Code, Cursor CLI, OpenCode, Antigravity CLI and raw API adapters behind one typed interface. Read-only turns resume the vendor's own native session, and write turns land as inspectable patches or complete file manifests.
How does Claudexor handle multiple subscriptions?
Since version 2.1 it has credential profiles, which are named Antigravity, Claude, Codex and Cursor subscription bindings side by side, each with Claudexor-scoped state and platform-declared credential custody. Live quota tracking covers Antigravity, Claude and Codex, and an opt-in policy rotates a spent account out of the way on typed vendor limits. Cursor has no usage source yet.
How do I install Claudexor?
Run npm install -g claudexor, which installs the claudexor and claudexord binaries, then run claudexor doctor. On macOS there is a signed and notarized DMG whose app bundles its own daemon, and the CLI and daemon also run on Linux with Node 20.19 or newer.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/razzant-claudexor)Star history chart for your README
[](https://hysenlabs.com/projects/razzant-claudexor)Was this useful?