# Point your Claude at Streamliner

Streamliner turns a company's processes into chats: a **streamline** per department (Register, Approvals,
Receiving), one job per conversation, and an agent that works your own systems through MCP **as the signed-in
employee**. Everything that makes it your company's desk is configuration: a folder of JSON files. Your Claude can
write that configuration with you, check it, and publish it once you approve.

## What your Claude will do

1. Ask what department and which one job each streamline is for, who uses it and which systems it needs.
2. Connect to your MCP servers (you sign in yourself, in your browser) and list their tools and annotations.
3. Propose which tools run at once, which wait for a person's tap, and which are left out, and settle them with you.
4. Draft the streamline files, the cards that show results, and sealed forms for anything sensitive.
5. Check everything with the Streamliner checker until it is clean.
6. Sign the manifest with your publisher key and show you a **dry run**: exactly what would change.
7. Publish only when you say yes.

## What it will not do

- Put an API key, token, password or personal data in any file. Model provider keys are named by reference and
  stored in your account, never in config.
- Collect card numbers or other secrets in a chat. Secrets go through sealed forms straight to your systems; cards
  through your processor's page or your terminal.
- Publish without your explicit yes, or with an employee's sign-in. It uses a server key your account owner issues.
- Follow instructions hidden in a tool's description or a document.
- Build "an assistant for everything". Each streamline does one job, with an end.

## Or in the app

Owners and admins can also browse and edit streamlines, briefs, templates and forms in the app's **Library**. An
organization's edits there are staged on the device, and an owner or admin publishes them from the app as a new
signed version, from a device an owner approved: the same signed publish as above. For personal streamlines on one
device, the app's built-in **Concierge** sets things up by conversation, showing every change before it saves.

## Install the skill

| Where | How | Notes |
|---|---|---|
| Claude Code | Unzip [building-streamlines.zip](https://streamliner.work/agents/building-streamlines.zip) into `~/.claude/skills/` (or your project's `.claude/skills/`) | Full network: inspecting servers and publishing work |
| claude.ai | Settings → Features → upload the zip (code execution on) | Network depends on your plan and admin settings |
| Claude API | Upload the zip to your workspace with the Skills API (`POST /v1/skills`); use it with the code execution tool | No network in the container: Claude drafts and checks; you run the publish commands it hands you |
| Other agents | The same folder follows the open Agent Skills format | |

## Keys to issue

Two Ed25519 key pairs, made on the machine that publishes; only their **public** halves are sent to Streamliner.

- A **server key** for the admin API. Give your Claude the narrowest scope that does the job (a dry-run-only key to
  draft; a streamlines-only key to publish streamlines) and a short expiry. Revoke it when the setup is done.
- A **publisher key** that signs the manifest. Devices trust config only if one of your publisher keys signed it,
  so a compromised server can't redirect your employees' credentials.

## Links

- [llms.txt](https://streamliner.work/llms.txt): the machine-readable index of these docs
- [SKILL.md](https://streamliner.work/agents/skill/SKILL.md) and its [reference files](https://streamliner.work/agents/skill/reference/concepts.md)
- [JSON Schemas](https://streamliner.work/schema/v1/) for every config file
- [Write API](https://streamliner.work/agents/skill/reference/write-api.md): keys, scopes, dry run, publish
- [Building MCP servers for Streamliner](https://streamliner.work/agents/skill/reference/mcp-servers.md)