Provider-native usage
Claude usage-credit spend, rolling session and weekly limits, Opus usage when reported, reset times, and local numeric history. ChatGPT surfaces show only recognized counters that OpenAI exposes to the page.
Private usage and efficiency tools for Claude and ChatGPT.
Know your limits. Waste fewer tokens. Keep your work private.
COMPANION shows provider-reported usage when trustworthy numbers exist, then adds local tools that help make prompts, replies, and files more efficient. No account. No backend. No quiet upload happening behind the curtain.
Completely local
Open source
No telemetry
Manifest V3
Native numbers or nothing.
COMPANION renders a provider value when the provider exposes a recognized numeric field. When it does not, COMPANION says so. It does not turn a guess into a polished-looking fact.
The extension as it exists now
Useful beside the composer. Quiet everywhere else.
COMPANION follows supported Claude and ChatGPT web surfaces, adapts to light and dark themes, and keeps provider boundaries separate under one consistent interface.
Limit warnings that earn attention
Optional toolbar badges and desktop notices appear around trustworthy 85% and 95% thresholds, with plain-English pace context instead of constant alarm bells.
Controls that stay out of the way
Hide the in-page widget, individual Claude metrics, Lifejacket, notifications, or the persistent badge. The extension is configurable without becoming a cockpit.
Lifejacket Mode
A local safety net for long prompts and expensive replies.
Lifejacket is optional and reviewable. It does not silently rewrite or submit your work.
- 1
Shorten the user’s prompt
A bundled quantized MobileBERT LLMLingua-style model runs locally through CPU and WebAssembly. Protected code, URLs, quotations, numbers, structured rows, negation, obligations, and bounds are checked before a result is offered.
- 2
Ask for shorter answers
A separate visible instruction requests brevity while retaining necessary facts, steps, and caveats. It works with prompt compression on or off.
- 3
Convert files to Markdown
Supported PDF, Office, OpenDocument, RTF, CSV, HTML, Markdown, and text files are parsed locally in a no-network sandbox, then appended to the draft. OCR is not included.
Lifejacket preview
Optimized prompt
Compare the shorter version with the original before anything leaves the browser.
Please provide me with a very detailed explanation of Explain how provider-native usage limits differ from token estimates, while making sure that you include necessary caveats.
Compression is lossy. The original remains one click away.
Privacy by architecture
Your work does not become COMPANION’s data.
Completely local means there is no COMPANION service waiting to receive your prompts, replies, files, or model inputs.
Read the privacy policyNo COMPANION accountNo signup, identity profile, or cloud sync.
No developer backendNo COMPANION API or remote processing service.
No analytics or telemetryNo behavioral tracking, advertising, or usage reporting.
No prompt or reply storageText is not retained by the extension.
No file uploadSelected files are processed locally and not kept.
No broad browsing accessHost permissions are limited to supported Claude and ChatGPT origins.
Provider-native numeric fieldsValidated locally
COMPANION extensionLocal widget and alerts
Your browserLocal settings and numeric history
Open source
Private claims should be inspectable.
COMPANION’s source, security architecture, build process, model provenance, release checks, and issue tracker are public. The project is independent and unofficial, with no affiliation to Anthropic, OpenAI, Microsoft, Hugging Face, or the model author.
Questions worth asking
FAQ
Is COMPANION really completely local?
Yes. There is no COMPANION account, backend, analytics service, remote model download, or developer-side file processing. Provider pages still receive the prompt you explicitly choose to send, because that is how Claude or ChatGPT works.
Does it estimate missing provider limits?
No. COMPANION shows recognized native numeric fields when available and leaves unsupported values blank. Claude token figures derived from native dollar counters are labeled as ranges rather than exact counts.
Can Lifejacket change a prompt’s meaning?
Compression is inherently lossy. COMPANION applies conservative safeguards and always presents an editable preview with optimized, original, and cancel choices. Use the original whenever necessary meaning is missing.
Does it work in the native Codex desktop app?
No. A native desktop shell is not a Chrome extension host. COMPANION supports Codex-aware web routes and shared agentic usage that OpenAI exposes through ChatGPT web surfaces.
When will it be on the Chrome Web Store?
The public launch is planned very soon after review and final smoke testing. Until then, the source and verified release materials remain available on GitHub.
Finish the work before the limit finishes the session.
COMPANION is coming to the Chrome Web Store soon.