Integrations
Connect your own OpenAI or Anthropic API key
Naffes CRM lets a workspace admin plug in the company's own OpenAI or Anthropic key, and four helpers then run on it: the deal AI copilot, Suggest with AI for a deal's next step, AI breakdown of tasks and filling a form from a customer's conversation. Everything else AI-related stays on the platform gateway, whatever key you add.
Before you start
- Who can do it: saving, changing or removing the key and running Test connection need workspace-admin rights.
- Plan: the Marketplace is included from Starter up and is not available on the Free plan. See pricing.
- The key: create it in your provider account (the API keys page at OpenAI, or Keys in the Anthropic Console). The Provider docs link on the setup screen goes straight there.
- One provider is enough: if both keys are saved, the OpenAI key is used for every call and the Anthropic key sits idle.
Steps
Open the Marketplace
In the sidebar, under Tools, choose Marketplace. Search for OpenAI or Anthropic Claude, or filter the catalog by the AI category.
Open the credentials screen
Click the app card, then Configure OpenAI or Configure Anthropic Claude. Install leads to the same screen and does no sign-in with the provider; the fields below are the real connection.
Enter the key and model
Paste the key into API Key. For OpenAI, Organization ID is optional, and an empty Default Model means
gpt-4o-mini. For Anthropic, type a current model ID such asclaude-sonnet-4-6in Default Model; do not leave it empty.Save
Click Save. The app switches to Configured, and a Last updated date appears under the field. The key is never shown in the browser again: type a new one to replace it, or leave the field empty on later saves to keep it.
Run Test connection
Click Test connection. For OpenAI, Naffes reads the list of models your account can use and shows OpenAI — connection OK with the model count and the latency in milliseconds.
Try a real helper
Open any task and, in its Subtasks section, click AI breakdown. The confirmation names the provider that answered, for example "Added 4 subtasks via openai", which is the quickest proof that calls now go through your key.
What your key powers, and what it does not
| Helper | Where to find it | Key used |
|---|---|---|
| Deal AI Copilot: Summarize deal, Draft outreach email, Suggest next step, Risk analysis, Discount advice, Ask anything | The AI assistant button in the deal drawer header | Your key first, then the platform gateway |
| Suggest with AI | The Next step card in the deal drawer, when no step is scheduled | Your key first; needs the Pro plan |
| AI breakdown | Subtasks section of the task drawer | Your key first |
| Fill the form, in the Fill from the conversation panel | A form document with a linked contact | Your key first |
| Naffes Smart, AI panel summaries on the deal page, next best actions, record insights, bulk AI, import column mapping, workflow AI steps, inbox reply suggestions, voice-note and call transcription | Across the app | Platform gateway only |
The OpenAI app's description in the Marketplace mentions summaries and smart replies; the table above is what actually happens. For the full picture of Naffes AI and its limits, see the AI feature page, and to add AI to automations, see AI steps in workflows.
How Naffes picks a provider for each call
- The workspace's saved OpenAI key, if there is one.
- Otherwise, the workspace's saved Anthropic key.
- Otherwise, the platform AI gateway set up by the Naffes operator.
- If none is available, the helper shows "Add an OpenAI or Anthropic key under Marketplace first."
So these four helpers may already work without a key of your own when the platform gateway is on. A workspace key is about control: calls go through your company's account, on the model you choose, billed directly by the provider.
Troubleshooting
- "AI returned an empty answer": the call reached the provider but was rejected or came back without text. Usual causes are a wrong or unavailable model name, no remaining credit, a usage limit, or a key revoked in the provider console.
- Test passes but helpers fail: Test connection checks the key, not the model; the OpenAI test only lists models. Clear Default Model and save to fall back to
gpt-4o-mini, which OpenAI's deprecations page listed with no shutdown date in September 2026. If calls then work, the model name you typed was the problem. - Anthropic with no default model: an empty field makes Naffes request
claude-3-5-sonnet-latest, an alias for the Claude Sonnet 3.5 models Anthropic retired on October 28, 2025, so every call fails. Set an active ID such asclaude-sonnet-4-6, Anthropic's recommended replacement. - Anthropic test fails with a valid key: in the current build the test sends a one-token request to
claude-3-haiku-20240307, which Anthropic retired on April 20, 2026. You may see Anthropic Claude — test failed with an error about the model rather than the key. Check with AI breakdown on a task instead. - You saved an Anthropic key but the toast still says openai: an OpenAI key is also saved and wins. Open the OpenAI screen and click Disconnect to clear it.
- Pause syncing does not stop the key: that action on the installed apps list only changes the install status, and the key keeps being used. Click Disconnect on the setup screen to stop it.
- Suggest with AI says it needs Pro: that button depends on the Pro plan's AI feature even with your own key, while the deal copilot works from Starter because it belongs to the Marketplace.
Questions
Questions 4
Will my key make Naffes Smart use my chosen model?
No. Naffes Smart chat, deal summaries, record insights and the other core features run only through the platform gateway. Your key serves the four helpers in the table.
Who pays for calls made with my key?
Your OpenAI or Anthropic account, at that provider's prices. Watch spending in the provider's own console: the AI usage numbers shown in Naffes do not log every call and are not a cost monitor.
What data is sent to the provider?
The deal copilot sends the deal name, stage, pipeline, amount, win probability, close date, up to three contact and company names, the next open task and the rep's prompt. It does not send emails or notes. Fill from the conversation reads the contact's messages the rep is allowed to see, and nothing is saved until the rep reviews it.
Can teammates see the key after it is saved?
No. The secret field is never returned to the browser; the screen shows only that it is set and when, and most secrets are encrypted at rest. Non-secret fields such as the model name are not hidden in the same way.
Sources
- Anthropic: Model deprecationsAccessed
- OpenAI: DeprecationsAccessed