PipesHub filters by source permission, and its installer generates the secrets
PipesHub is an open-source platform for securely connecting enterprise knowledge to AI. Give AI agents trusted context and your team permission-aware search with verified citations across your business systems.
At a glance
- What is it?
- PipesHub is an Apache-2.0 platform that indexes company systems such as Slack, Google Drive, GitHub, Microsoft 365 and Notion, then answers questions with citations while enforcing per source access controls. It installs with one curl command that runs an interactive installer, generates a .env with randomised secrets, and can bring any model provider inside your own network.
- Who is it for?
- PipesHub suits an organisation that already has documents spread across several SaaS tools and needs an assistant to answer from them without leaking what a given person should not see. It does not suit a team that needs its own model hosted elsewhere, because the model is expected to run inside your own network or a managed cloud offering that is still on a waitlist.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- pipeshub-ai/pipeshub-ai4,121 ★ · +383
| Repository | 2026-09-11 | 2026-11-15 |
|---|---|---|
| pipeshub-ai/pipeshub-ai | 3.7k | 4.1k |
383 Stars in the last 65 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
One command installs a Compose stack, not an application
The deployment story begins with a piped shell command:
curl -fsSL https://get.pipeshub.com/install | bashThe requirement above it is Docker with Compose v2, and the script downloads the deployment files for the latest release into a directory called pipeshub before launching an interactive installer. When the installer finishes you open the application on localhost at port 3000. The point to hold on to is that this is a stack rather than a binary: the project also states that PipesHub can be run locally or deployed on any server using Docker Compose, and that the image published to Docker Hub is the normal path, since the one-command installer always uses prebuilt images. Building the images locally is a separate route that requires a cloned repository and an extra flag on the installer.
The installer checks your machine, then writes the secrets for you
The installer's own checklist is published, and it is the best description of what a deployment involves. It checks Docker, RAM and disk prerequisites. It asks whether you want a slim or a full deployment, and lets you customise the graph database, the message broker and the key value store. It generates randomised secrets and writes them into a .env file, then pulls the images and starts the stack, waits for PipesHub to become healthy, verifies that it is reachable and prints the URL. Two flags change that behaviour: a yes option accepts all defaults and skips the interactive prompts, which is described as friendly to continuous integration, and a version flag pins a specific image tag, with 0.7.0 given as the example.
A plain HTTP deployment gives you a white screen
One failure mode is documented in enough detail to be worth quoting. On a cloud server the project asks for an HTTPS endpoint, because browsers block certain requests over plain HTTP, and it names Cloudflare, Nginx and Traefik as ways to terminate TLS. The symptom is stated plainly: a white screen after an HTTP only deployment is typically caused by that restriction. Which means the default local experience, opening localhost on port 3000, works fine, and the same image deployed to a public host with a plain endpoint does not. That is a deployment detail rather than a bug, but it is the one most likely to consume an afternoon, and it is fixed by a reverse proxy rather than by a configuration value inside PipesHub.
Permissions are enforced per source, not filtered after the answer
The trust story has two halves. Answers come with precise block level citations back to the original documents, so a claim can be traced to the passage it came from. Access control is applied at the source level, so a user only sees what they are authorised to see, and the results a search returns are filtered by what each person is allowed to see. The example the README works through shows the difference: a question about why retry logic in the billing worker changed, answered from the incident postmortem, the pull request, the chat thread and the design document, each cited, and returned only if the asker is allowed to see them. The path to that is a personal access token that needs no administrator, then one command for Claude Code or one configuration file for Cursor or Codex, which the README puts at about ten minutes of work once the data is indexed.
The model runs inside your network, and the cloud is still a waitlist
Deployment of the model is the third choice the platform asks you to make, and the framing is deliberate. The bring your own model item says any large language model provider can be used, deployed inside your own VPC, with the claim that data never leaves your infrastructure. That is the difference between a hosted assistant and an on premises one, and it is why the feature list keeps self hosting as a headline rather than as a setting. The managed alternative is named separately: PipesHub Cloud is described as coming soon, with a cloud waitlist for early access. So at present the fully managed path is not something you can switch to, and the self hosted path is the one the installer is built for.
Fifty connectors, two indexing modes, and a knowledge graph underneath
The connectivity claim is a count. Connectors are described as fifty or more, with real time and scheduled indexing available out of the box, and the systems named in the overview are Slack, Google Drive, GitHub, Microsoft 365 and Notion. Real time versus scheduled is the distinction that matters operationally, since one is a webhook and the other is a crawl. Above the connectors sit two retrieval layers. Knowledge graph retrieval is described as graph backed and capturing relationships across enterprise data, which is the answer to questions that span systems rather than one document. Multimodal support covers image, diagram and scanned file understanding, plus voice based interaction, so the pipeline is not limited to text files.
Four SDK surfaces for agents, and a showcase template for what you build
The distribution badges in the header describe the client surfaces: an npm package for the SDK, a PyPI package named pipeshub-sdk, a separate repository for a Go SDK, and an npm package published under the MCP name. The MCP client is the path for editor assistants, which is how the README frames giving Claude Code, Cursor or Codex access to company knowledge. For people writing their own retrieval rather than wiring an assistant, the SDK starter and a search example live in a separate examples repository, and the project asks that anything built with them be submitted through a showcase issue template. That request is a small but telling design decision: the reference implementations are not in this repository, so the core stays free of demonstration code.
The image is layered, and the published base has known gaps
The Dockerfile explains its own structure in comments, which is unusual and useful. Slow layers, covering the system packages, Rust, Python dependencies, machine learning models and the runtime stack, live in a separate base file, and both base images are overridable through build arguments so a registry mirror can be substituted when the published images are missing. The slim application image deliberately omits the pre-baked model, which makes it about 1.3 GB smaller and moves the model download to first use. The build reconciles the base image against the current Python manifest so packages added since the base was published still land in the application image, and it installs a crawler setup step plus a Chromium browser for Playwright. Two known gaps are patched inline: fallback CJK fonts for LibreOffice, because the published runtime base included Writer and Calc only, leaving presentation files unreadable.
Editorial conclusion
PipesHub suits an organisation that already has documents spread across several SaaS tools and needs an assistant to answer from them without leaking what a given person should not see. It does not suit a team that needs its own model hosted elsewhere, because the model is expected to run inside your own network or a managed cloud offering that is still on a waitlist. Before deploying, put TLS in front of it with Nginx, Traefik or Cloudflare, since a plain HTTP deployment shows a white screen, and read the install script first if your policy requires reviewing a piped shell command.
Frequently asked questions
How do I install PipesHub?
With Docker and Compose v2 installed, run curl -fsSL https://get.pipeshub.com/install | bash. It downloads the deployment files for the latest release into a pipeshub directory and launches an interactive installer, and you then open the application on localhost at port 3000.
Can I review the PipesHub install script before running it?
Yes. The README gives the alternative of downloading it to a file, pipeshub-install.sh, reviewing it with a pager such as less, and only then running it with bash. The piped form is the shortcut version of the same script.
How does PipesHub handle permissions for search results?
Access control is enforced at the source level, so users only see what they are authorised to see, and answers carry precise block level citations to the original documents. A personal access token can be minted without an administrator.
Why does my PipesHub deployment show a white screen?
Because it is being served over plain HTTP and browsers block certain requests in that case. The project asks for an HTTPS endpoint on a cloud server, with Cloudflare, Nginx or Traefik used to terminate TLS, and names the white screen as the typical symptom.
Which systems does PipesHub connect to?
Fifty or more connectors, with Slack, Google Drive, GitHub, Microsoft 365 and Notion named as examples, and with real time and scheduled indexing available out of the box. A separate examples repository holds a starter and a search example for building on the SDKs.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/pipeshub-ai-pipeshub-ai)Star history chart for your README
[](https://hysenlabs.com/projects/pipeshub-ai-pipeshub-ai)Was this useful?