Gadak

Find the thread in your backlog.

_{English · 한국어 · 日本語}

Query your Jira backlog with SQL.

gadak mirrors the Jira Cloud projects and Confluence spaces you select (issues, comments, history, wiki pages) into one SQLite file on your machine, indexed together and searchable with no network. Use the mirror through the desktop app, a browser tab (gadak serve), the CLI, or MCP. Writes go to Jira first. One binary, no gadak account. It is 0.x software from one maintainer; status and promises are below. Atlassian's own Rovo MCP server is hosted and needs nothing installed, but it has no aggregation tool and no offline reads; the trade is in alternatives.

▶  Open the live demo  ·  534 issues, in your browser, right now.
Changelog  ·  what shipped.

Query Jira with SQL

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"

JQL has no GROUP BY. This query counts unresolved issues in each epic; the REST API returns the rows and leaves the counting to you. Datasette Lite runs this query on the demo snapshot in your browser with nothing installed. More questions JQL cannot ask, as SQL: docs/RECIPES.md.

Measured 2026-08-26 against a live Cloud site (3,296 issues; medians, CLI startup included):

The method, the re-measurement history, and the rows where gadak loses (the first full sync, the watch tick on a quiet site, one sync interval of staleness) are in docs/BENCHMARKS.md.

Install

macOS app, CLI included:

brew install --cask midagedev/tap/gadak

CLI only, on macOS or Linux. The same UI opens in a browser tab via gadak serve:

brew install midagedev/tap/gadak-cli

Connect to Jira Cloud, then open the address gadak serve prints (http://gadak.localhost:7777). The first sync runs inside serve, newest issues first, and the list fills while it runs:

gadak init && gadak serve         # mirror a Jira Cloud site
gadak init --local && gadak serve # no Jira — start on the tracker gadak ships with

A site needs one API token, a user token created with no scopes; it covers Jira and Confluence on the same site, and the mirror sees what your account sees. You pick what it mirrors: --projects for Jira, --spaces for the wiki, which stays off until you name spaces. Jira Server and Data Center connect with gadak init --server and a Personal Access Token; the wiki stays off there (no Confluence Server client).

gadak init --projects ENG,PROD --spaces ENG

Windows: the desktop app is on the Microsoft Store. The Store signs it, so neither SmartScreen nor Smart App Control objects, and a Store install puts the gadak command on PATH as well (since 0.20.2). For the CLI without the Store, take gadak_<version>_windows_amd64.zip (or arm64) from the latest release, unzip, put gadak.exe on PATH. The release's desktop zip (Gadak-<version>-windows-x64.zip) stays unsigned, and a SmartScreen block there is a missing signature, not a virus finding (why). If it blocks, install from the Store, and do not turn Smart App Control off.

The window follows the browser or OS language (English, Korean or Japanese; Settings switches it). The signed dmg, the Linux tarball, Docker, upgrades: docs/INSTALL.md.

Other workspaces. gadak --workspace <new> migrate --from <old> moves a synced workspace onto another origin; add --to linear to target a Linear team. A second machine pairs with a home serve: gadak --workspace laptop init --pairing-code-stdin.

Data and network

• How much is copied. The Cloud projects and spaces you named, as your account sees them; gadak adds no elevation and no service account.
• What stays local, and how fresh. One SQLite file. The first full sync is the slow part: 3.7 minutes for the benchmark site (3,514 issues and 462 pages). After that gadak serve syncs every 60 seconds by default (syncIntervalSec), plus an hourly reconcile that removes issues you can no longer see. Reads are one interval behind Jira.
• Where the token is. ~/.gadak/config.json, mode 0600. Credentials never reach SQLite, a log, or a snapshot.
• What leaves the machine. No telemetry. gadak talks only to what you configured; SECURITY.md is the complete list and docs/NETWORK.md walks every connection and its off switch. Four reads still ask Jira: viewing an attachment, gadak issue --editmeta, gadak fields, and gadak api (a passthrough).
• What a write does. It goes to Jira first; the mirror refreshes after Jira accepts. A write Jira rejects fails then and there. Nothing is queued.
• What changes with an agent. It sends what it reads to its model; see Agents.

The mirror is a cache you can throw away. Delete the directory and nothing is lost.

Agents

Let your coding agent query the same mirror through the CLI or MCP. Reference: docs/MIRROR.md; one paste per host: docs/AGENT_SETUP.md.

gadak skill install

That copies one SKILL.md for Claude Code and starts no process; the agent runs short-lived gadak commands. Name another host to install the same file there: gadak skill install codex, and the same for agents, cursor, gemini, opencode and grok. For Claude Desktop, which has no shell, register the same mirror as an MCP server (gadak mcp install claude is the Claude Code registration; Claude Desktop never sees it):

gadak mcp install claude-desktop

_{The shell is in the window (⌘K → Terminal, or Ctrl+`). gadak claim binds the pane to the issue, and a Claude Code session started in it filters the list, then saves and opens a dashboard in the same window. Nothing but the two prompts is scripted; the stretches where the agent is working are time-lapsed. The Korean and Japanese READMEs carry the same take recorded in their own language. Recorded from e2e/demo/terminal-claude-demo.spec.ts via record-terminal-claude.sh.}

Two rules carry most of the value. Filter on status_category and priority_rank rather than a display name: Jira translates those per account, so priority = High is silently zero rows on a Korean-language site. And open the issues a query returns in the gadak window: gadak sql --no-header "…" | gadak views open --keys - puts the agent's answer on your screen, and gadak views open --jql '…' lands pasted JQL as chips.

Writes (create, edit, comment, transition, claim, link, and the wiki's page verbs) go through the origin before the mirror refreshes. On Jira and Linear, an agent's comments and the issues it creates carry its name (off switch: gadak config set actor.trailer false). On the built-in tracker, the person at the machine has the same byline: gadak me set "Your Name" authors their own writes as them.