Cloudflare AI gateway

Optional. Runs inside your own Cloudflare account so the AI credential never touches SlowBooks' company file: SlowBooks holds only a shared secret that opens your one Worker, and the Worker calls Cloudflare's Workers AI models with the account's own binding.

If you would rather not run a Worker, Settings → AI Insights also takes a Cloudflare API token directly (provider Cloudflare Workers AI (direct)), or a key for Anthropic, xAI, Groq, OpenAI, Google Gemini or any OpenAI-compatible endpoint. This page is for the Worker.

Why bother?

• Your keys stay in Cloudflare. SlowBooks only ever holds a shared secret scoped to your one Worker. Even if someone copies the company file, they can't talk to Workers AI as you.
• Per-installation lockdown. Every office installs its own Worker in its own Cloudflare account. One compromised install can't reach another, and one abused install can't burn someone else's allowance.
• Zero leakage. The Worker never logs the shared secret, never echoes it back, and never stores request bodies. The secret is compared in constant time.
• Cost. Workers AI has a free daily allowance (10,000 neurons a day at the time of writing) and paid accounts are billed per neuron. An AI Insights run is capped at 1,024 output tokens, which on the default 70B model is a few dozen runs a day on the free allowance; Cloudflare's pricing page has the current figures.

One-time setup (about 10 minutes)

1. Cloudflare account

The free plan is enough: https://dash.cloudflare.com/sign-up

2. Install wrangler

Wrangler, Cloudflare's command-line tool, needs Node.js 22 or later (Ubuntu 24.04's packaged Node is 18, which it refuses). Install Node from https://nodejs.org or your package manager's current channel, then, from the cloudflare/ folder of your SlowBooks checkout:

cd cloudflare
npm install wrangler@latest
npx wrangler login

wrangler login opens a browser tab; approve the request and come back to the terminal. The commands below use npx wrangler; a global install (npm install -g wrangler) works too, as plain wrangler.

3. Generate a shared secret

openssl rand -hex 32

Copy the 64-character string. You'll paste it twice: once into Cloudflare (step 4), once into SlowBooks (step 6). Never commit it.

4. Store the secret in Cloudflare

npx wrangler secret put AUTH_TOKEN
# Paste the hex string from step 3, press Enter

Wrangler encrypts the secret at rest inside Cloudflare. It never shows up in source control, in logs, or in the Worker's responses.

5. Deploy the Worker

npx wrangler deploy

Plain deploy: this wrangler.toml has no environments, so --env production would create a second Worker named slowbooks-ai-production with no secret, which SlowBooks never calls. Wrangler prints your Worker URL, e.g. https://slowbooks-ai.yourname.workers.dev. Copy it.

6. Wire SlowBooks to the Worker

1. Open SlowBooks → Settings → AI Insights (an administrator's setting).
2. Provider: Cloudflare Worker Gateway (self-hosted).
3. Worker URL: the one wrangler deploy printed.
4. API key: the shared secret from step 3.
5. Model: pick one from the list. The list is this Worker's allowlist (ALLOWED_MODELS, below); the default is @cf/meta/llama-3.3-70b-instruct-fp8-fast.
6. Click Save AI settings, then Test. A good answer names the model and says the provider answered.

No account ID is needed for the gateway; that field belongs to the direct provider.

Smoke test from the command line

curl -sS https://slowbooks-ai.yourname.workers.dev/v1/models \
  -H "Authorization: Bearer <your-shared-secret>"

curl -sS https://slowbooks-ai.yourname.workers.dev/v1/chat/completions \
  -H "Authorization: Bearer <your-shared-secret>" \
  -H "Content-Type: application/json" \
  -d '{
        "model": "@cf/meta/llama-3.3-70b-instruct-fp8-fast",
        "messages": [
          {"role": "user", "content": "Reply with the word ok and nothing else."}
        ],
        "max_tokens": 16
      }'

The first answers the allowed models in OpenAI's list shape; the second an OpenAI-shaped response with choices[0].message.content: "ok".

Environment variables

All optional except AUTH_TOKEN. Set the optional ones in the [vars] section of wrangler.toml and deploy again; a value set in the Cloudflare dashboard is overwritten by the next deploy unless wrangler.toml has keep_vars = true.

AUTH_TOKEN (required)

The shared secret SlowBooks sends as a Bearer token. Set it with npx wrangler secret put AUTH_TOKEN, never in wrangler.toml.

ALLOWED_MODELS (optional)

Comma-separated model IDs the Worker will run, and the list GET /v1/models answers (what SlowBooks' model picker shows). Default:

@cf/meta/llama-3.3-70b-instruct-fp8-fast,@cf/mistral/mistral-7b-instruct-v0.2-lora

Cloudflare retires models from time to time (@cf/meta/llama-3.1-8b-instruct, in the old default, went on 2026-05-30); the current list is at https://developers.cloudflare.com/workers-ai/models/. Pick text-generation models.

DEFAULT_MODEL (optional)

The model used when a request names none. Must be in ALLOWED_MODELS. Default @cf/meta/llama-3.3-70b-instruct-fp8-fast. SlowBooks always names the model it was given, so this only matters for other callers.

LOG_VERBOSE (optional)

Set to "1" to log each request's model, message count and token usage (never the text or the secret) to the Worker's logs, visible with npx wrangler tail. Default "0".

ALLOWED_ORIGINS (optional)

Browser origins allowed for CORS. SlowBooks calls the Worker from its own server, not from the browser, so this is only for other clients of yours. Unset, no CORS headers are sent.

Rotating the shared secret

npx wrangler secret put AUTH_TOKEN   # paste a fresh `openssl rand -hex 32`

Then paste the new value into SlowBooks → Settings → AI Insights → API key and Save AI settings. Anything still using the old secret gets 401 at once.

Operational notes

• Live logs: npx wrangler tail streams every request to your terminal (the Authorization header is never logged).
• Custom domain: bind the Worker to a domain you own from the Worker's settings in the Cloudflare dashboard, then put that address in SlowBooks.
• Limits: the Worker refuses bodies over 512 KB, more than 64 messages, and models not on the allowlist, and clamps max_tokens to 4,096. The [limits] section of wrangler.toml caps CPU time per request (paid plans); request rate limiting is Cloudflare's rate-limiting binding, not covered here.
• Auditing: every invocation is counted on the Worker's page in the Cloudflare dashboard, under its metrics and logs.

What this does not protect against

• A compromised SlowBooks install still has the shared secret, so it can still invoke your Worker. Rotate the secret if you suspect compromise; unexpected traffic shows in the Worker's metrics.
• Workers AI is a third-party service. Cloudflare's terms and privacy policy apply to what you send through it, as with every AI provider SlowBooks offers.

Uninstalling

npx wrangler delete

That removes the Worker. SlowBooks' AI features then fail with a 502 until you redeploy or choose another provider under Settings → AI Insights.