omni-dev

An intelligent Git commit message toolkit with AI-powered contextual intelligence.

Transform messy commit histories into professional, conventional commit formats with project-aware suggestions.

What it does

AI-powered improvement — Claude analyses diffs and suggests meaningful commit messages.
Contextual awareness — understands project structure, conventions, and prior work.
Safe amendments — improve single or batched commit messages with working-directory validation.
PR creation — generate complete pull request descriptions in one command.
Atlassian integration — read and write Jira issues and Confluence pages, with JFM ↔ ADF conversion.
Browser bridge — drive authenticated HTTP requests through a logged-in browser tab without exfiltrating its cookies or tokens.
Media transcripts — fetch YouTube captions as SRT, VTT, TXT, or JSON, with language fallback and optional ASR/translation.
Conversation history — export your Claude Code sessions to versionable JSONL and Markdown for work-logs and review.
MCP server — drop omni-dev's capabilities into Claude Code via the bundled MCP binary.

Install

CLI only:

cargo install omni-dev

CLI plus the omni-dev-mcp server (for Claude Desktop, Claude Code, Cursor, etc.):

cargo install omni-dev --features mcp

Or with Nix:

nix profile install github:rust-works/omni-dev

Demo

Seven short asciicasts — each paired with the artefact it produces. The source for every cast lives at rust-works/omni-dev-demo.

A bug-report JFM file exercises nested numbered/bulleted lists, a warning panel, a task list with mixed states, dates and status pills, an inline code mark, a fenced JavaScript code block, a pipe table, and an expand block. One CLI call creates the issue in Jira; jira read rolls it back to identical Markdown.

Resulting Jira issue page rendered in Atlassian Cloud — Resulting Jira issue — panel, dates, status pills, nested lists, task list, code block, table, and expand all rendered from the JFM source.

Three commits ahead of main. omni-dev git branch create pr reads the commits, asks Claude for a PR title and body that summarises them, pushes the branch, and opens the PR via gh.

Resulting GitHub pull request page — Resulting PR on rust-works/omni-dev-demo#1 — title and body generated from the three commits on the branch.

Three deliberately messy commits (wip, stuff, fix). omni-dev commit message twiddle reads each diff, asks Claude for a Conventional Commits message that reflects the change, and rewrites the branch via git rebase.

GitHub commit list showing the rewritten commit messages — Top three commits on demo/twiddle after force-push — the three messy messages are gone.

omni-dev browser bridge serve runs a local server that drives fetch() calls inside an authenticated browser tab — so requests carry the tab's SSO session, but the cookies and tokens never leave the browser. Here it pulls a Loki datasource and its labels straight out of a logged-in Grafana, with no API keys to manage. A confused deputy by design; the security model is the load-bearing part.

grafana.internal — DevTools › Console

> (async () => { /* omni-dev bridge snippet */ })()
< Promise {<pending>}
[omni-bridge] connected to ws://127.0.0.1:9999
[omni-bridge] authenticated · origin https://grafana.internal
[omni-bridge] GET /api/datasources → 200 (1 datasource)
[omni-bridge] GET …/loki/api/v1/labels → 200 (4 labels)

The pasted snippet runs in-page, so every request inherits the tab's grafana_session cookie and passes CSRF — while the bridge process only ever sees the response bodies, never the credentials.

omni-dev transcript youtube pulls captions straight from a video. info and list-langs inspect what is available; fetch renders a track as srt, vtt, txt, or json. Here it inspects the first video ever uploaded to YouTube and prints its transcript — --auto falls through to ASR captions, --translate synthesises a missing language.

me-at-the-zoo.srt

1
00:00:00,000 --> 00:00:05,800
All right, so here we are in front of the elephants.

2
00:00:05,800 --> 00:00:11,200
The cool thing about these guys is that they have
really, really, really long trunks.

3
00:00:11,200 --> 00:00:19,000
And that's cool. And that's pretty much all there is to say.

Same transcript, written to disk as WebVTT/SRT with -o me-at-the-zoo.srt — ready to attach to a video or feed to a search index. Drop --format json for cue objects with millisecond timing.

omni-dev ai claude history sync exports your Claude Code sessions — prompts, responses, thinking, and tool calls — to a target directory as one .jsonl (and/or rendered .md) per chat. Re-runs skip unchanged sessions by mtime, so it is safe to schedule; --since, --project, and --dry-run scope the run, and --prune removes exports whose source is gone.

~/hist — after sync

-Users-me-omni-dev/
  0f1a…2b.jsonl   # append-only event log
  0f1a…2b.md      # rendered transcript
-Users-me-spike/
  a1b2…d4.jsonl
  a1b2…d4.md

── 0f1a…2b.md ──────────────────
## 👤 User
Refactor the transcript fetcher to support VTT.

## 🤖 Assistant
Here's the plan: extract a Format enum…

One folder per project (the encoded cwd slug), one file per session UUID. The JSONL is append-only and byte-identical across runs; the Markdown is a readable transcript for work-logs and review.