gadak
Query your Jira backlog with SQL.
gadak mirrors the Jira Cloud projects and Confluence spaces you choose into one SQLite file on your machine. Issue titles, bodies, comments and wiki pages land in one search index, and searching it opens no connection. SQL over the same file answers the counts JQL cannot express. The desktop app, the browser tab, the CLI and a coding agent read that file; writes go to Jira first, and the file is a cache you can delete.
Why not Rovo MCP?
Rovo MCP is hosted by Atlassian and needs no local MCP server installation. It searches Jira and Confluence and has write tools, but no native aggregation tool and no offline reads. gadak runs SQL against a synchronized local mirror, so a count over the whole backlog or a join across the change history is one query. The costs are a local binary, an initial sync, and reads that trail Jira by one sync interval.
| Rovo MCP (hosted) | gadak (local mirror) | |
|---|---|---|
| Where the query runs | Atlassian’s cloud | your machine, in SQLite |
| Open issues per epic | no native aggregation tool | one GROUP BY; 22 ms on 3,296 issues (2026-08-26) |
| Network required for search | yes | no, once the selected data is synced |
| Searches Jira and Confluence | yes | yes, for the spaces you select |
| Writes: comment, transition, assign | yes | yes, through Jira first |
| Freshness | Atlassian’s hosted data | the last synced state, one interval behind |
| Local setup | none | a local binary and an initial sync |
| Rate limits | 429 and a Retry-After | none on reads — they come off your disk |
| Who maintains it | Atlassian | one person, and Apache-2.0 |
The same comparison for jira-cli, Linear and Jira’s own UI: docs/FAQ.md#how-it-compares
JQL has no GROUP BY.
The REST API returns issues a page at a time, and no totals. Once the data is a local file, the count is one query:
gadak sql "select epic_key, count(*) from issues_full where resolved_at is null
and epic_key <> '' group by epic_key order by 2 desc" Run this query on the demo snapshot in your browser (Datasette Lite, nothing installed) →
On a live Cloud site with 3,296 issues (2026-08-26, medians): 4,761 ms over 8 API pages aggregated client-side, against 22 ms for this query.
The same questions, measured
| Jira REST API | gadak | ||
|---|---|---|---|
| Simple filter, 100 issues | 583 ms | 19 ms | 31× |
| One issue + full changelog | 710 ms | 28 ms | 25× |
| Free-text search | 543 ms | 41 ms | 13× |
| Open issues per epic (GROUP BY) | 4,761 ms | 22 ms | 214× |
| A count over the change history | no native aggregate | 14 ms | — |
| Rate limit on mirror reads | 429 + Retry-After | none | — |
Measured 2026-08-26 against a live Atlassian Cloud site: a real work project, 3,296 issues. Medians. The gadak numbers include full CLI process startup. In this measurement the epic count took 8 API pages, aggregated client-side, against one query on the mirror. The history count has no native aggregate; crawling and counting it client-side takes about 28 minutes. The rate-limit row is about reads, which stay local. A first full sync of that site took 3.7 minutes (2026-09-09), and the mirror trails Jira by one sync interval. Method, re-measurements, and the rows where gadak loses: docs/BENCHMARKS.md.
Use the CLI from your coding agent
You and your agent read the same file. The CLI is the agent interface: create, claim, transition, and SQL over the mirror, while the board in front of you updates. Writes go through Jira first, and an agent’s comments and the issues it creates carry its name. An agent that reads your mirror sends what it reads to whatever model it talks to; scope the mirror to what the agent should see.
Install the skill for Claude Code:
gadak skill installFor Claude Desktop, register the MCP server:
gadak mcp install claude-desktop One paste per host → docs/AGENT_SETUP.md
More recordings: dashboards, a team theme, a launcher, a live MCP session → docs/SHOWCASE.md
The same cache, on a phone Beta
The iOS app pairs with a gadak serve you already run, over your own network — no gadak account, and nothing hosted in between. Whatever the cache holds, the phone shows: fields, attachments, linked issues, wiki pages, the active sprint and every sprint behind it. What is one line to say, the phone writes through the origin first, the way the desk does — a comment, a label, a due date, a photo, an issue filed under its epic. Editing a page or a custom field, authoring a view, laying out a dashboard and moving cards stay on the desktop, and the app says so where the control would be.
Beta, and that is the whole status: builds go to TestFlight for internal testers, and there is no public download. An App Store listing waits on a way for a reviewer to see the app without joining your network.
What gets copied, where it lives, and what leaves your machine
- Which Jira. Atlassian Cloud with an API token, or Jira Server / Data Center with
gadak init --serverand a personal access token. One Cloud token covers Jira and Confluence on the same site; Confluence Server has no client, so a Server workspace mirrors issues only. - How much is copied. The projects you name with
--projects, as your account already sees them. The wiki stays off until you name spaces with--spaces. - What stays here, and how fresh. One SQLite file on your machine. A first full sync, then one sync interval behind Jira. Delete it and sync again; nothing is lost.
- What leaves this machine. No telemetry, no analytics, no gadak account. gadak opens only the connections you configured, and reads come out of the file. Four verbs still ask the origin: viewing an attachment,
gadak issue --editmeta,gadak fieldsandgadak api. - Where the token is.
~/.gadak/config.json, mode0600, sent only to your own site. It never reaches SQLite, a log, or a snapshot. - What a write does. It goes to Jira first, and the cache refreshes after Jira accepts. A write Jira rejects fails then and there; nothing is queued locally.
- What changes with an agent. An agent that reads the cache sends what it reads to its own model. gadak sends nothing. Scope the cache to what the agent should see.
Every outbound destination and its condition → SECURITY.md
Every connection and its off switch → docs/NETWORK.md
Install
The desktop app, CLI included:
brew install --cask midagedev/tap/gadakCLI only, on macOS or Linux. The same UI opens in a browser tab via gadak serve:
brew install midagedev/tap/gadak-cliOn Windows, the desktop app is on the Microsoft Store.
Connect to your team’s Jira Cloud site. It asks for the site, your email, an API token and the projects to mirror; add --spaces to include the wiki:
gadak init && gadak serveStatus and limits
- Status: 0.24, still 0.x. Sync, reads, write-through, desktop, web, CLI and MCP are verified against a live site.
- One maintainer, currently. Apache-2.0. The mirror is ordinary SQLite: if gadak went away tomorrow, the file opens in any SQLite client and your issues are still in it.
- Keep sprint planning, administration, page editing in a UI, and anything that cannot tolerate a sync interval of delay in Jira.
- Four origins, one set of verbs: Atlassian Cloud, Jira Server / Data Center, Linear, and the built-in tracker. What each one refuses is in one table, measured cell by cell.
What each origin supports · Earlier projects in this area · Hard questions · Who maintains this · Source
If you used it on your own project, say so
What question did gadak answer, and did you go back to it?
Keep real issue data, tokens, and site URLs out of public reports.