Please do not open a public issue for a security problem. Use GitHub's private vulnerability reporting — it goes only to the maintainers.
Please include what you were running (version, OS, Python), what you did, and what happened. You'll get a first response within a week, and credit in the release notes when a fix ships unless you'd rather stay anonymous.
Supported: the latest release on master. There are no backported security fixes for
older versions while the project is pre-1.0.
Taskuary is a local-first, single-user application. It binds to 127.0.0.1 and
assumes the only person who can reach that port is the person sitting at the machine.
That assumption is the whole security model — the notes below are what follows from it.
- Every request needs a token. On first start Taskuary writes an owner token
(
[server].token) and a separate agent token to~/.taskuary/config.toml, and the API refuses a request without one (a provider's sign-in redirect and a vendor's signed push prove themselves another way). The browser gets the owner token from a one-time launch key when Taskuary opens it (taskuary --linkprints a new one); a page opened without the key carries no token. Agent sessions carry the agent token, which the routes that send, approve or change settings refuse./api/healthstays open on purpose so a container HEALTHCHECK can pulse the process without the token. - Don't expose it to a network you don't trust. If you must reach it from elsewhere,
send the token as the
X-Taskuary-Tokenheader (TASKUARY_TOKENsets it in Docker) and put it behind TLS you control. The token gate is a lock on a door, not a hardened perimeter. - Docker still assumes localhost.
docker composepublishes127.0.0.1:7787; the process inside listens on0.0.0.0only so Docker can reach it. Changing the publish bind to0.0.0.0:7787(LAN) is the same as a local install with[server].host = "0.0.0.0". - The API executes things by design. Reports run SQL and PowerShell,
/api/tools/runruns a connector's query or script, agent dispatch runs your configured CLI, and the terminal opens a real shell — all with your credentials, on your machine. Anyone who can reach the API can do those things./api/tools/runstill refuses a catalog connection that is off or not marked tool (cards are seeded on first launch, so "never connected" is not the same as "no row"). A report written by an AI agent can also reach the web, so on Claude Code it reads only the folder it runs in and the files its instruction names - never Taskuary's data folder; Codex's read-only mode cannot be kept to a list. - Credentials are stored locally. Connector secrets in
~/.taskuary/taskuary.dbare sealed with Windows DPAPI or the system keychain where one exists, and stay plain text where neither does;config.toml(the tokens among them) is plain text. They are write-only through the UI (never returned to the browser). Sealing stops a copy of the file carrying the keys, not a program running as you: the files are only as protected as your user account and disk encryption. - Agents act with your permissions. A CLI agent configured with an auto-approve flag will edit files, run commands, and push to repos without asking. Scope its PAT and working directories to what you actually want it to touch.
- The text Taskuary reads is written by other people, and some of it is written for the AI (prompt injection). Triage runs with no tools, no agent works a stranger's task until you press Trust, and nothing is sent without your press. An agent that is running still acts as you. See Security for what is and is not covered.
- AI triage sends message content to whichever AI you connect (an API provider, or your local CLI agent). Nothing leaves the machine until you connect one.
- Give GitHub PATs the narrowest repository access and permission set that works.
- Keep
~/.taskuary/out of backups that sync somewhere you don't control. - Attachment URLs only serve files under
~/.taskuary/attachments/— aPaththat escaped that folder is a 404. SVG/HTML always download; they are not rendered as a document on the app origin. - Read
~/.taskuary/taskuary.logand the in-app audit log (Settings → Audit integrity — it's a hash chain) if you want to know what ran and when.