What doesn't work yet, or works with a catch, in the current release. Each entry says what to do meanwhile. Format-specific limits are also on each format's page. Fixed issues leave this list and appear in the changelog; security problems are published as security advisories.
Last reviewed: 2 November 2026.
1. Installation and clusters
| Issue | Workaround |
|---|---|
| Depotex can't be served under a path (https://tools.example.com/depotex/) | Give it its own host name |
| Nothing stops a second server from using its own local disk as storage: two servers with separate disks each serve only what was uploaded to them | In a cluster, use object storage or one shared volume mounted at the same path on every server |
| Some limits are read from each server's environment only (listed under Environment variables); servers with different values answer differently | Give every server the same environment |
| pool_max_conns added to DATABASE_URL after a server's first start is ignored: the URL saved in its bootstrap file wins | Set it at first start, or remove the server's bootstrap file and start it again with DATABASE_URL set |
| A license change of tier reaches some features only after a restart | Restart the servers (one at a time in a cluster) after applying a license with another tier |
| Settings → Clustering at 1440 px wide with the settings menu open hides the Node column's name | Widen the window or collapse the settings menu |
| A Redis failover (Sentinel or a cluster shard) can lose the last writes, so two servers may briefly hold the same lock | Harmless: PostgreSQL keeps the records correct |
| During a rolling upgrade, live messages over 7.9 KB can't take the PostgreSQL fallback while Redis is down | Keep Redis up during upgrades |
2. Sign-in and identity
| Issue | Workaround |
|---|---|
| Signing out of Depotex doesn't end the session at the identity provider (no single logout) | Sign out at the identity provider too on shared machines |
| Turning on Require single sign-on doesn't end password sessions already open, and OAuth refresh tokens issued before it keep working | Revoke sessions (Settings → Access → Users) and applications' tokens after turning it on |
| The setup wizard takes an 8-character first password, though people's passwords need 12 by default; there is no breached-password check | Choose a long first password; enforce password quality in your identity provider |
| Anyone who knows an account's name can lock it for AUTH_LOCKOUT_DURATION by failing sign-ins, and the answer reveals the account exists | The per-address limit caps it; unlock in Settings → Access → Users; use single sign-on |
| API tokens keep working while their owner must change a temporary password | Revoke the person's tokens when resetting their password |
| SCIM: no bulk operations, sorting or ETags; filters only attribute eq "value" | Enough for Okta and Entra ID provisioning |
| Entra ID and Google Workspace sync read the whole directory each run | Fine for tens of thousands of people hourly; sync less often for larger tenants |
| Global roles are fixed (member, organization admin, super admin): you can't define your own | Use groups and space roles for fine-grained access |
| Emails are in English and German only, and show the app name but not your logo | — |
| Invitation and password reset links keep their one-time token in the address | They expire (AUTH_INVITATION_TTL, AUTH_PASSWORD_RESET_TTL) and work once |
3. Repositories and formats
| Issue | Workaround |
|---|---|
| Raw and Maven uploads need a Content-Length (chunked uploads are refused with 411) | Clients send it by default; for scripted uploads, upload a file rather than a stream |
| Uploading from the browser: hosted repositories only; not Docker, Terraform, Hugging Face, Git LFS or Conan; files wait whole on the server's temporary disk | Use the format's own client for those |
| Proxies in pass-through mode don't verify checksums, and the first download of a big file waits for the whole fetch | Keep files (cache) in proxies where checksums matter |
| Quarantine doesn't check pass-through proxies, and holds whole versions, not single files | Use caching proxies under quarantine rules |
| The block list covers Maven, npm, Docker (tags and digests, not layers), raw and the formats whose pages say so | See Block list |
| NuGet: v3 only (no v2 / OData feed), so nuget.exe 2.x and PowerShellGet 2 can't install | Use the .NET CLI, NuGet 3+ or PSResourceGet |
| Not supported by design: Docker manifest schema 1 (Docker 1.9 and older), Composer 1, git clone of Hugging Face repositories, OCI Helm charts in Helm repositories | Use current clients; put OCI charts in a Docker repository |
| npm: no group repositories | Use scope registries |
| Repository groups: write rules don't apply to npm and Docker; pip search through a group searches the first proxy only | See Repository groups |
| Rebuild metadata: APT, Cargo and Conan report missing files but can't rebuild them; Hugging Face has nothing to rebuild from | Restore the files from a backup of the storage |
| People with only a repository role don't see its space in the spaces list | Share the repository's link |
| Cleanup: one rule set per policy; only the last run is recorded | Use several policies |
4. Security data
| Issue | Workaround |
|---|---|
| SBOM coverage differs by format (version ranges in libraries aren't resolved; Helm, Terraform, Git LFS and raw read only SBOM files they contain) | Upload SBOMs from CI for what you deploy; see SBOMs |
| Inspection and image scans run one at a time per cluster; a version a policy would refuse can be downloaded until the queue reaches it | Quarantine new versions (Enterprise) so they wait for inspection |
| Vulnerability matching is by OSV and package URL only; VEX per document only; share links can't be revoked | — |
5. Plugins and AI agents
| Issue | Workaround |
|---|---|
| AI clients can't register themselves; an administrator registers each one | The MCP plugin's setup page fills in the registration |
| MCP write tools can only quarantine and release; document search is keyword-only; MCP keys can't be rotated or extended | Create a new key before the old one expires |
| Plugins are enabled for the whole installation, not per organization | — |
6. Operations
| Issue | Workaround |
|---|---|
| Prometheus metrics cover rate limits and Redis health only, not HTTP latency, database pool or jobs | Measure latency at your proxy or ingress |
| The metrics listener (METRICS_ADDR) has no authentication of its own | Bind it to a private interface or use METRICS_TOKEN |
| Backups don't include artifacts, nor plugin schemas in PLUGIN_DATABASE_URL; the wizard's restore finds backups only in the storage's import folder or an upload | Protect artifact storage with your storage platform; back up a separate plugin database yourself |
| Erasing a person (GDPR) covers the database; backups and external logs keep data until they expire | Include backups and log retention in your erasure process |
| A server uses old storage settings for up to a minute after a change | Change storage when it's quiet |
| Live alerts need /api/ws routed to the server; without it they arrive every 30 seconds | Route it (see Install Depotex) |