Local Docker demo (no instance yet)

This is the playbook when there is no running PipesHub. Offer a local Docker install. Do not start it until they agree. Do not invent a RAG stack while you wait. If they already have an instance, go back to For coding agents and skip to Connect.

Honest cost

On the order of 30–60 minutes (image pull + first health wait, which the installer budgets at up to 420 seconds)
A 16 GB-class machine (installer floor is 15000 MB). WSL: 10 GB. Docker Desktop on macOS: at least 8 GB allocated to the VM
They still complete first-run in a browser: first account + an LLM. Vendor API key goes in the onboarding form (or a password manager), never pasted into a conversation. On a laptop, Ollama is enough — no vendor key. Embedding can use the system default. Search and chat both fail until an LLM is configured (HTTP 500: LLM configuration is missing). That is not an empty index.
Slack / Drive / Jira still need a browser later (OAuth). For a laptop demo, upload files to the Knowledge Base (no OAuth, no extra Docker mounts) or use Local FS with a bind-mount

--yes skips installer prompts, including the low-RAM abort. Check RAM before invoking. Keep the stack on localhost until first-run is done — POST /api/v1/org is unauthenticated first-run, so whoever reaches a fresh instance first owns it. --yes without PIPESHUB_PROJECT always targets Compose project pipeshub-ai. If that stack already exists, --yes updates it — it does not create a second copy. Never attach to someone else’s corpus.

Step	Agent	Human
1. Preconditions	`docker info`, RAM	Agree to install
2. Install	`install --yes` (new project if a stack exists)	—
3. Health	Poll `health/services`, not `/health`	—
4. First-run	Wait; poll `org/exists`	Account, LLM, PAT/OAuth in the UI
5. Data	Point at KB upload or Local FS mount	Drop files / pick a folder
6. MCP	Write config to localhost	—
7. Prove	`pipeshub_search` with backoff	—

1. Preconditions

docker info >/dev/null   # must succeed
# Linux / WSL:
awk '/MemTotal/ {print int($2/1024)}' /proc/meminfo   # MB; need ≥15000, or ≥10240 on WSL

If Docker is missing or RAM is below the floor, stop. Do not run --yes anyway. If docker ps already shows a PipesHub (pipeshub-ai, port 3000, or another project), do not run a bare --yes. Use a separate project and port (below) or attach to the instance they already have (Connect).

2. Install (slim)

Empty machine (no existing PipesHub):

curl -fsSL https://get.pipeshub.com/install | bash -s -- --yes

That writes ./pipeshub/ (or PIPESHUB_DIR) and defaults to port 3000, project pipeshub-ai. This machine already has a PipesHub — required, not optional. Export the knobs before the pipe so both curl and bash see them (a prefix on curl does not reach the installer):

export PIPESHUB_DIR="$PWD/pipeshub-demo"
export PIPESHUB_PROJECT=pipeshub-demo
export PIPESHUB_PORT=3200
export PIPESHUB_DEPLOY_TYPE=slim
curl -fsSL https://get.pipeshub.com/install | bash -s -- --yes

From a clone, ./install.sh --yes at the repository root is a wrapper; compose files live under deployment/docker-compose/. Same PIPESHUB_PROJECT / PIPESHUB_PORT rules apply. More flags: Quickstart and Advanced deployment. Leave FRONTEND_PUBLIC_URL blank (--yes does). Do not expose port 3000/3200 on a public interface until first-run is finished.

3. Wait until healthy

Poll GET http://localhost:<APP_PORT>/api/v1/health/services (default 3000, or the port you set). Ready when query, connector, indexing, and docling are "healthy". Do not wait for embedding. Do not poll /health — that path returns the SPA and is 200 immediately.

PORT="${PIPESHUB_PORT:-3000}"
for i in $(seq 1 84); do   # 84 × 5s ≈ 420s
  json=$(curl -sf "http://localhost:${PORT}/api/v1/health/services" || true)
  echo "$json" | python3 -c "
import json,sys
s=(json.load(sys.stdin).get('services') or {})
need=('query','connector','indexing','docling')
sys.exit(0 if all(s.get(k)=='healthy' for k in need) else 1)
" && echo ready && break
  sleep 5
done

4. Human first-run (browser)

They open http://localhost:<APP_PORT>:

Create the first account (Individual is enough).
Onboarding: configure an LLM. Paste a vendor API key into the form (not into chat), or pick Ollama: a model they already have installed (ollama list), endpoint left at the default http://host.docker.internal:11434. Compose already maps that hostname to the host (docker-compose.yml extra_hosts). Confirm from the install directory:
docker compose -p "$PIPESHUB_PROJECT" exec pipeshub-ai \ curl -sf http://host.docker.internal:11434/api/tags
Embedding: system default is fine for a demo.
Developer Settings: create an OAuth app or a personal access token. They put the PAT in a local env file or the client’s secret store — never paste it into chat.

Poll until the account exists:

curl -sf "http://localhost:${PORT}/api/v1/org/exists"
# {"exists":true}  → first account exists. MCP and search are NOT ready yet.

Do not treat GET /api/v1/org/onboarding-status as readiness. It only gates the UI wizard (notConfigured → /onboarding). pipeshub_search returning HTTP 500 with LLM configuration is missing means they skipped the LLM step.

5. Index something that does not need OAuth

Preferred for a Docker demo: Knowledge Base upload in the UI (no extra mounts). Have them drop a few .md / .txt / PDF files. Local FS (docs): no login, but the path must exist inside the container. After install, in the directory the installer printed (./pipeshub or PIPESHUB_DIR), add docker-compose.override.yml and recreate:

services:
  pipeshub-ai:
    volumes:
      - /absolute/path/to/their/docs:/data/demo:ro

cd "$PIPESHUB_DIR"   # or ./pipeshub
docker compose -p "$PIPESHUB_PROJECT" --env-file .env \
  -f docker-compose.yml -f docker-compose.override.yml up -d

Then in Workspace Settings → Your Connectors → Local FS, set Local folder to /data/demo, enable sync, run a sync. A host path that is not mounted is invisible (docker compose exec pipeshub-ai ls /data/demo must list files). Do not start Slack/Drive OAuth unless they asked for those systems.

6. Attach MCP

Cursor and Claude Code can use http://localhost:<APP_PORT>/mcp. Claude.ai and Claude Desktop’s custom connector cannot (they need public HTTPS). Config is in Connect. For a PAT, prefer the stdio bridge with env interpolation — do not commit the token.

7. Prove it works (do not skip)

Call pipeshub_sources, then pipeshub_search on a term from the files they uploaded. Back off (e.g. 15s, 30s, 60s, up to several minutes):

Result	Meaning
No sources / empty KB	They have not uploaded or connected a folder. Not a product failure.
HTTP 500 mentioning `LLM configuration is missing`	They have not finished the LLM form (vendor key or Ollama). Not an empty index.
HTTP 401 on `/mcp`	Token missing or wrong instance.
Sources present, `hits` empty	Still indexing. Wait and retry. Do not report “PipesHub is broken.”
At least one hit with `recordId`	Demo working.

Do not curl first-run APIs (POST /org, login, PAT create, LLM config) from a chat. PAT create returns the secret in JSON. Use the bootstrap script instead: First-run without a browser. Then put the skill + AGENTS.md in their repo — For coding agents.

Welcome To PipesHub

System Overview

Authentication

Mail Configuration

AI Providers

Connectors

Integrations

Agents

Toolsets

MCP Server

User Management

Deployment

Developer

Additional Resources

Honest cost

1. Preconditions

2. Install (slim)

3. Wait until healthy

4. Human first-run (browser)

5. Index something that does not need OAuth

6. Attach MCP

7. Prove it works (do not skip)

​Honest cost

​1. Preconditions

​2. Install (slim)

​3. Wait until healthy

​4. Human first-run (browser)

​5. Index something that does not need OAuth

​6. Attach MCP

​7. Prove it works (do not skip)

Honest cost

1. Preconditions

2. Install (slim)

3. Wait until healthy

4. Human first-run (browser)

5. Index something that does not need OAuth

6. Attach MCP

7. Prove it works (do not skip)