Is this just a paraphraser that will quietly invent my opinions?

That is the failure mode it is built against. The source supplies meaning and your references supply style only; facts, requests, stance, certainty and commitments must stay intact, and reference anecdotes must not turn into invented personal experiences. Every mode retains your original if the candidate fails a meaning check or the tweet-length check, and each evaluated version exposes its raw answers and probabilities.

Will it post to X for me?

No. Nothing is automatically posted, and the extension does not trigger X's own Post or Reply controls. Applying a suggestion goes through Chrome's native insertText editing command, the result is verified, and if the editor rejects the operation the card asks you to check your draft again.

What permissions does the extension actually ask for?

Its only persistent host permission is loopback HTTP to 127.0.0.1, and network calls are restricted to a single local port. It attaches only to X's post and reply editors, explicitly excluding direct messages, settings and sign-in surfaces, and it does not request permanent access to X. API keys, X tokens and the CSRF token stay in the backend and the content script cannot supply arbitrary fetch URLs.

Which model does which job?

An OpenAI model runs the clarity pass and Jev runs the style passes and meaning checks, with optional overrides for the writer and judge model names. In AI refine plus personalize, a clarity pass is followed by up to two TypeSafe-guided style passes; personalize only skips the clarity pass; AI refine only makes one clarity pass plus one Jev call with the two meaning checks and no style scoring.

How do I run Mimicry locally?

Requires Python 3.12+ and uv. Run uv sync --extra dev, add OPENAI_API_KEY and TYPESAFE_API_KEY to .env, then uv run mimicry-web and open 127.0.0.1:2719.

What does each of the three writing modes execute?

AI refine plus Personalize runs a Luna clarity pass then up to two TypeSafe-guided style passes; Personalize only skips the clarity pass; AI refine only runs one Luna pass and one TypeSafe call with two meaning checks.

Which X permissions does the extension request?

Only tweet.read and users.read, with no publishing or direct-message integration. Its only persistent host permission is loopback HTTP on port 2719, not X.

CommunityJev Tools & CLIs

Mimicry

A Python workspace plus Manifest V3 Chrome extension that turns a rough opinion into a draft in your own voice, with an OpenAI clarity pass and Jev as the style judge, and per-phrase accept or dismiss.

See highlights

Open github.com

Stars: 0
Language: Python
Published by: Community
Author: jxucoder
Last verified: Oct 2026

jxucoder/mimicry

User repository on GitHub

View on GitHub

Rewrite AI drafts in your own voice with a bounded TypeSafe feedback loop.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Turn a rough opinion into a clear draft, then match it to your own past writing with less empty framing and repeated explanation. Browse the real X feed with the Chrome side panel, or use the standalone local workspace. The source supplies meaning; your references supply style only. Facts, requests, stance, certainty, and commitments must stay intact. Reference anecdotes must not become invented personal experiences. Requires Python 3.12+ and uv. From the project root:

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Meaning is locked, style is borrowed: facts, requests, stance, certainty and commitments stay intact, and anecdotes must not be invented.
The result separates the AI rewrite from the version in your voice, with a diff view for removed and added text.
Three modes run different provider sequences; AI refine only skips style scoring and does not use your personal samples.
Accept replaces one underlined range via Chrome's native insertText command and verifies the result. Nothing is posted automatically.

Watch out

You need two providers, not one: OPENAI_API_KEY and TYPESAFE_API_KEY in a local .env, with your reference text and draft sent to both, and run artifacts written under an ignored runs/ directory. The shipped example voices and draft are original fictional samples, not human benchmark data, so the demo tells you nothing about quality on your writing. This is an unpacked development extension you load yourself, not a Web Store release, and the README is candid that X's controlled-editor behaviour still has to be checked in your target Chrome profile because the DOM selectors can change when X updates its app. For reply drafts you tick a visible checkbox to include the parent post; it never infers context from other feed posts. One further limit stated plainly: refine-only meaning probabilities are not a claim that clarity improved. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jevclip

Video tool where Jev scores subtitle segments and code assembles a short highlights cut plus a de-watered full cut with timestamped summaries.