Skills
How to work.
Your team's checklists, formats and procedures as versioned instructions. Share one with people, groups or everyone; publish a new version and every chat and agent uses it.
The AI system your IT runsApache-2.0 core
One AI system for your whole organization.
Chat for everyone, agents in Chrome and Office, SDKs for your developers, and the models on your own GPUs. What large companies need whole departments for, your IT runs as one system.
$ pipx install fadenstack
$ faden deployOne Linux server with Docker. Your GPU machines join from the console.
- People and appschat · Chrome · Office · your apps
- Fadenstackone endpoint · skills · knowledge · tools · policies
- Your GPU machinesyour models · outside providers only where you allow them
- Built in Germany
- Runs on your hardware
- OpenAI-compatible /v1
- Console in EN · DE · FR · ES · 中文
- Apache-2.0 core
Why one system
What large companies need whole departments for.
One team runs the GPUs, another the gateway, another the chat, and yet another decides which data may go where. Most organizations don't have those teams. Fadenstack gives your IT all of it as one system, on your own servers, run from one console.
| Area | What organizations usually end up with | With Fadenstack |
|---|---|---|
| Chat | Chatbot accounts, bought team by team | One chat for everyone, on models you choose |
| Agents | A different assistant in every app, each with its own rules | Agents in Chrome and Office that follow the same rules as the chat |
| Apps | Every project wires up its own API keys | One endpoint and one SDK for every app |
| Know-how | Prompts copied between documents and chats | Shared, versioned skills, with a record of which one was used |
| Personal data | A policy document, and hope | Redacted by your policy before a model sees it |
| Models and GPUs | A GPU server someone set up once | Machines, clusters and models, run from the console |
| Oversight | Separate bills and no audit trail | Usage, cost and audit per request, in one place |
What each part of your organization gets
Three kinds of people, one system.
For everyone
A chat assistant on your own models.
It works like the chat assistants people already know, with your documents, your team's shared skills and the tools IT has connected.
- Projects, files and a history that stays private
- Type
$to use a skill your team wrote - On phones too, in five languages
For developers
One endpoint and an SDK for agents.
An OpenAI-compatible /v1, plus open-source SDKs for .NET and TypeScript that add sessions, tools and a ready chat panel. Every app gets the organization's skills, knowledge and tools through that one endpoint, under the same rules as the chat.
- OpenAI clients work for chat, models and embeddings: change the address and the key
- Agents bring their own tools, which run on the user's side
- Docs and samples on developers.fadenstack.com
from openai import OpenAI
client = OpenAI(
base_url="https://ai.example.internal/v1",
api_key="<your Fadenstack key>",
)
reply = client.chat.completions.create(
model="team-assistant",
messages=[{"role": "user",
"content": "Summarise ticket 4471"}],
)
print(reply.choices[0].message.content)
For IT
One console for all of it.
Users and roles, models and GPU machines, skills, MCP servers, privacy rules, usage per user and the audit log, in one place, on your servers.
- Add an NVIDIA machine with one line and an approval code
- Pick a model; the marketplace says whether it fits
- See where every answer was computed
Agents
Fadenstack where people already work.
Assistants in the browser and in Office, free to use. They answer with your organization's models, through your Fadenstack server, and they ask before they change anything.
The browser agent
Reads the page you have open, answers about it, reads PDFs, compares tabs and writes into the form or mail you are working on.
Agents in Office
Summarise a document, add a chart, comment a draft. By default every change waits for the user's go, and most can be undone.
The mail agent
Reads the thread, drafts the reply and leaves it for you to send. It never sends anything itself.
Set up once, by IT
Which model answers, who may use the agent and which of the app's tools it may run.
Conversations stay on the device
The server keeps the rules and the usage; the chat history stays in the app.
A desktop agent In development
The same assistant outside the browser and Office.
Managed once
Skills, knowledge and tools, for every chat and agent.
Every chat and every agent draws on the same three things, and IT manages each of them in one place.
Knowledge
What it knows.
Your documents, searched by the model itself, with the source named in the answer.
Tools
What it can do.
MCP servers are registered once on the server, hosted ones or your own, and every agent that is allowed can call them.
on the serveron the user's sideboth
Agents can also bring tools of their own, like the open document or files on the user's machine. Each session's policy decides where tools run.
How a request travels
Follow the thread of every request.
Chats, agents and apps all take the same path through Fadenstack, and every request is recorded: who sent it, which model answered, where it ran and which skills and tools it used.
- 1
Limits & routing
Limits apply and the model's name finds its route.
- 2
Context
History, the team's skills and the allowed tools join the request.
- 3
Privacy
Where your policy says so, personal data is redacted before it reaches a model.
- 4
Model
Your machines, your network, or a provider you have allowed.
- 5
Tools
On the server or on the user's side; your privacy policy covers their results too.
- 6
Answer & audit
The answer streams back; the audit log records what was used.
You typeDraft a reply to Alice Meyer about invoice 4471.
The model seesDraft a reply to <PERSON> about invoice 4471.
Data stays yours
Know where every answer was computed.
Fadenstack sorts every model by where it runs: your machines, your private network, or an outside provider. The dashboard shows the split; the trace shows each request.
- Outside providers only where you have added and allowed them
- HTTPS by default, and mutual TLS between the server and your machines
- Built in Germany, running on your own servers
AI trafficWhere requests were served
Last 7 days
- Your machines82%
- Private network13%
- Outside providers5%
- Unknown0%
Watch
See it before you install it.
Short videos for each kind of reader: the whole idea in ninety seconds, a day with Fadenstack, and the steps from a bare server to the first answer.
With the release
Length 1:30
For everyone
Fadenstack in 90 seconds
With the release
Length 3:00
Users and buyers
A day with Fadenstack
With the release
Length 7:00
Operators
From install to first answer
Get started
From a bare server to the first answer.
One Linux server runs Fadenstack, and it needs no GPU. The machines with GPUs join from the console.
$ pipx install fadenstack
$ faden deploy
✓ Docker and Compose v2 found
✓ Configuration written # random keys and database passwords
✓ HTTPS with a certificate for this server
✓ Services started
Console https://ai.example.internal
API https://ai.example.internal/v1# the line from Machines → Add a machine
$ curl -fsSLO https://ai.example.internal/api/install/faden-agent.sh \
&& sudo sh faden-agent.sh --join https://ai.example.internal
✓ NVIDIA GPU found
✓ Agent running as a service
… Waiting for approval. Code on this machine: K7Q-4TD
✓ Approved in the console. Joined as gpu-01Search Hugging Face › chat models
an-open-chat-model
✓ Fits your cluster: 2 machines, 2 GPUs
Settings filled in for this model
Offered as team-assistant
[ Host ]# the chat
https://ai.example.internal
# the agents: Chrome, Word, Excel, PowerPoint, Outlook
Server https://ai.example.internal
# your apps: any OpenAI client
base_url "https://ai.example.internal/v1"Documentation
Help for the people who use it, and docs for the people who build on it.
help.fadenstack.com
Help
Also built into every install at /help, for the version you run.
- Use: the chat, skills, the browser and Office agents
- Administer: users, models, agents, privacy rules, usage
- Operate: install, upgrade, machines, HTTPS, backups
developers.fadenstack.com
Developers
Everything you need to build on Fadenstack.
- The OpenAI-compatible API, with its reference
- The .NET and TypeScript SDKs: agents, tools, the chat panel
- MCP servers and the skills format
Editions
Open source at the core, free agents on top.
Apache-2.0
Community
Free
The whole server, open source.
- Chat, console and the
/v1API - Skills, knowledge and MCP servers
- Personal-data redaction and the residency view
- GPU machines, clusters and models
- Roles, governance rules and the audit log
Free to use
Agents
Free
For every Fadenstack server.
- The browser agent for Chrome
- Word, Excel, PowerPoint and Outlook on Windows
- The desktop agent In development
- Open-source SDKs to build your own
Commercial licence
Enterprise Planned
On request
For regulated teams, on the same platform.
- Single sign-on
- Content policies on prompts, answers and tool calls
- Limits on what each agent's tools may do
- An auditor's view with exports
- Support with an SLA
Planned features are what we are building next; their scope and timing may change.
Compare the editions
from the front
Why the name
der rote Faden
In German, der rote Faden (the red thread) is the guiding idea that runs through something and holds it together as a whole.
The expression goes back to Goethe's Die Wahlverwandtschaften (Elective Affinities, 1809). He describes a red thread woven through every rope of the English Royal Navy: it could not be pulled out without unravelling the rope, and even a small piece could still be recognized as belonging to the Crown.
Fadenstack is that thread through your organization's AI. Chat assistants, agents, data, models, tools and infrastructure are connected once, governed in one place, and traceable end to end.
Run your organization's AI as one system.
Install on one Linux server, add a GPU machine from the console, and invite your people.