Upgrade notes#
- Back up the data volume before you upgrade. The first start runs a database migration automatically. To roll back, restore that backup instead of starting the older image on the new data.
- The
latestimage tag now points to 2.0.0. To stay on 1.3.9, pin the image to1or1.3. The Helm chart and the Compose files uselatestby default, so pinimage.tagto control when your pods move to 2.0.0. - Members, Usage and all settings are now under Workspace Settings, with one page per section at
/admin/<section>. Old/settings/<section>,/members,/usageand?tab=links redirect to the new pages. - The settings of a repository are now at
/projects/<owner>/<repo>/-/settings. - Members who sign in with SSO can be disabled, but not removed, in Context7. Remove them in your identity provider.
- Only administrators can refresh Google Drive and Notion libraries, even when anonymous refresh is allowed.
- Only administrators can import libraries. Scripts that import with a member’s API key now get a 403 response.
- The GitLab token is now sent only to
gitlab.comand to the self-hosted GitLab host set in Workspace Settings › GitLab. Before, any host with “gitlab” in its name got it. On the first start, Context7 sets that host from your existing GitLab libraries when they all use one host. If they use several, the server log names them; set the right one in Workspace Settings › GitLab. After the upgrade, check the host in Workspace Settings › GitLab, and clear it if it is not your GitLab server. - The first syncs after the upgrade can use more language model tokens: large Confluence pages that were skipped for their size are now indexed. Turning on the decision model, or leaving the Bedrock model empty in an Asia-Pacific region, also extracts each page again once.
Added#
- Added member management: administrators can change a member’s role, disable or enable sign-in, and remove members who sign in with a password. Context7 does not let you demote, disable or remove the last active administrator.
- Added support for SAML 2.0 single sign-on for on-premise installations.
- Added Google Drive, Google Docs and Google Slides sources that index Google Docs and Google Slides and sync them daily. Administrators pick Shared Drives, folders or single files in a tree that loads folders as you open them and searches all of Drive.
- Added three ways to connect Google: a service account key, with optional domain-wide delegation; the server’s own Google credentials, such as Workload Identity; or a dedicated Workspace user who signs in once through an internal OAuth app, with no key and no sharing outside your domain.
- Added a Notion connector that indexes approved pages and database rows and syncs them daily.
- Added automatic descriptions and language detection for Google Drive and Notion libraries when the administrator leaves them empty.
- Added an optional decision model for page classification, duplicate checks and search reranking. It works with OpenRouter, OpenAI, Vercel AI Gateway or a System One-compatible endpoint, and the language model decides whenever it fails.
- Added the source document to informational snippets in query results, so agents can cite and link the original page.
Changed#
- Redesigned the on-premise web interface. One navigation bar holds Dashboard, Repositories and MCP Test, a status button for the parse queue and system checks, and a help menu. Members, Usage and all settings are grouped under Workspace Settings, with Google, Notion and the Git hosts in an Integrations group. A new parse queue page shows running, waiting and recent parses. A new repository settings page lets administrators change the branch, folders and exclusions of a repository.
- Updated the on-premise service image to run on Node.js 24 instead of Node.js 22.
- Improved the accuracy of APIDOC context by grounding endpoints within their respective page sections and dropping irrelevant snippets.
- Replaced the agent runtime that on-premise CodeDocs uses to write documentation. CodeDocs keeps the same page IDs when it regenerates identical documentation, so a regeneration does not repeat the extraction.
- Large Google Drive, Notion and Confluence pages are now indexed in full instead of being skipped for their size.
- Updated the AI provider model suggestions to each provider’s current models. The Gemini default is now
gemini-3.8-flash, and Bedrock suggests one inference profile per Claude model for the region’s geography. - Library pages show a branch only for Git repositories.
- The parse queue shows members only the libraries they can access, and deleting a library also deletes its parse history.
Fixed#
- Fixed parser stalls that occurred when processing very large pages or long exclusion lists.
- Improved the stability of the on-premise CodeDocs agent by isolating its process and applying timeouts to SDK requests.
- Fixed an issue where large HTML conversion inputs could exhaust worker resources.
- Fixed issues with context7.json folder traversal and case-insensitive enterprise path guards.
- Fixed an issue that could lead to non-linear snippet token counting when processing looped AI outputs.
- Fixed content that Google Drive, Notion and Confluence syncs could drop without a warning: short slides and pages are indexed, a failed page classification is retried, and Confluence pages picked by URL are indexed under the right library.
- Fixed Google Drive picker selections that did not match what the administrator checked, for example when unchecking a file inside a large folder or checking a folder while it loads. Failed loads now show a Retry button.
- Fixed a cancelled first Google Drive sync that left the library without its selection, so Refresh can restart it.
- Fixed Google Drive syncs failing on a temporary Google token error. Context7 now retries it for all three connection methods.
- Fixed Notion libraries that stopped syncing without an error after the Notion token was replaced with another integration. They now show a sync error that tells you to add them again.
- Fixed Bedrock in Asia-Pacific regions suggesting inference profiles that current models do not have.
- Fixed pages that went blank when they failed to load. They now show a loading state and a reload prompt.
Security#
- Hardened security by binding GitLab tokens to exact hosts and restricting library import functionality to administrators only.
- Improved data privacy by masking secrets in on-premise queries before they are sent to Context7 Cloud.
- Enhanced data privacy by processing requests for restricted local libraries entirely on-premise, preventing their identifiers from being sent to Context7 Cloud.
- Restricted refreshing Google Drive and Notion libraries to administrators, because a refresh reads content through the connector’s credentials.
- A Google reconnect now refuses a sign-in as a different Google account, so the connected identity cannot change by mistake.
Added#
- Added native AWS Bedrock support for Anthropic Claude inference, Amazon Titan embeddings, reranking, and source-code documentation.
Changed#
- Made AWS Bedrock model defaults region-aware.
- Improved parser extraction cache reliability through versioning.
Fixed#
- Fixed multiple onboarding issues, including license key requirements, wizard defaults, GitLab token errors, and detailed parse status reporting.
- Fixed several small issues in the setup flow.
No user-facing changes were recorded for this release.
Fixed#
- Fixed custom model configuration for generated documentation indexing.
Fixed#
- Fixed custom model configuration for code indexing.
Added#
- Added an official Helm chart for easier on-premise deployment.
- You can now bound and cancel inference requests.
Changed#
- Enabled running the on-premise image as a non-root user.
Fixed#
- Provided more detailed reasons when a parse extracts no content.
- Fixed an issue with source code indexing on the on-premise image.
Added#
- You can now cancel in-flight and queued parse jobs.
- Added support for parsing SCP-style SSH repository URLs.
Changed#
- UI actions now accurately reflect user roles and permissions.
- Improved website crawler coverage.
Fixed#
- Ensured vector index integrity during parsing and querying.
- Improved database resilience by preventing a single database issue from affecting all replicas.
- Automatically dropped projects with invalid URLs to prevent them from getting stuck.
Security#
- Improved security by redacting Git tokens from logs, errors, and support bundles.
- Enhanced security by requiring administrator privileges for certain API configuration test endpoints after initial setup.
- Made support bundle generation admin-only by default.
Added#
- Added library access control: restrict which libraries each user can query over the MCP server and the REST API, scoped to their single sign-on groups or to individual users, with a workspace-wide allow-by-default or restrict-by-default posture.
- Added namespace access rules that govern every library under a path, with rules on an individual library taking precedence over the longest matching namespace.
- Added an access audit log that records access-rule changes and blocked access attempts.
- Added support for running the on-premise image as multiple replicas behind a load balancer, backed by a shared PostgreSQL database and a shared vector store.
Changed#
- Improved reranking for queries with several requirements, and reduced duplicate snippets in responses.
Added#
- Added the option to create a Confluence source by pasting page URLs, optionally including their sub-pages, instead of browsing the page tree of an entire space.
Added#
- Added the option to configure the Other Git SSH deploy key and its known hosts from the dashboard, instead of mounting a key into the container.
Added#
- Added an Other Git integration for ingesting private repositories from self-managed Git hosts such as Gerrit, Gitea, and self-hosted Bitbucket, over HTTPS with stored credentials or over SSH with a deploy key.
Added#
- Improved the relevance of responses from the MCP server and API through new reranking capabilities.
- Added support for GitOps-based routing for Confluence and other on-premise data parsers.
Added#
- Added a REST API for programmatic library import.
Added#
- Added support for ingesting content from Confluence Cloud and Data Center.
- Added full end-to-end integration with GitHub Enterprise Server.
- Added support for ingesting content from llms.txt files.
Fixed#
- Improved parsing robustness by preventing process hangs and cleaning up orphaned staging data.
Fixed#
- Fixed the path used for GitHub Apps on GitHub Enterprise Server.
Fixed#
- Fixed the GitHub App client to correctly target the on-premise GitHub Enterprise Server API.
Added#
- GitOps declarative repository management: define the repositories to index in a Git manifest (
repos.yaml) and have Context7 reconcile them on startup, on a schedule, on manual trigger, and on GitHub push webhooks. Supports per-repo branch, folder, and file filters, plus pruning of repositories removed from the manifest. - Generic OIDC single sign-on for the on-premise dashboard.
- Configure a GitHub Enterprise Server host from the dashboard so the GitHub App and tokens target your own server instead of github.com.
Added#
- Initial tracked release of the Context7 On-Premise image.