A complete studio for designing, testing, and operating agents on the Meta Business Agent Platform. Skills, knowledge, connectors, evals, live chat, webhooks, all in one workspace.
git clone https://github.com/v-senthil/whatsapp-business-agent-studio.gitnpm install && npm run devThis feature is not yet released for general availability. To enable it for your WhatsApp Business Account, an admin must open Meta Business Manager and accept the WhatsApp Business Agent terms and conditions.
https://business.facebook.com/latest/whatsapp_manager/business_ai?business_id={Business ID}&asset_id={WABA ID}Replace {Business ID} and {WABA ID} with your own IDs. If the page loads and prompts you to accept the T&C, your WABA is eligible for the Beta. After signing in, the studio also generates this link for each of your WABAs on the home page.
Every surface of the Meta Business Agent Platform is here, designed to feel native, not like a raw API reference.
Author skills as prompts, wire connectors to your APIs, and expose them as tools the agent can call. Everything is versioned in JSON so you can export a config and reproduce a workspace anywhere.
Upload files, register websites, curate FAQs. Import and export CSV for anything list-shaped.
Rich WhatsApp components the agent can send: carousels, CTA buttons, published Flows, images, interactive lists, and location prompts. Each with a trigger instruction and an enable toggle.
Cap token usage over rolling 1, 7, or 14 day windows at the Business Manager level. Multiple caps stack; whichever is reached first takes effect.
Run Meta eval cases with live status polling. Save any test-chat conversation as a scenario, then replay it side-by-side with the current config, per-turn match / differs / error badges.
Talk to your agent inside the studio, with sticky conversation IDs and one-click save-as-scenario.
Signed webhook capture with a live SSE feed, filter bar, and full request inspection.
Drop in a doc, get FAQs, skills, and connector suggestions you can apply with one click.
Scoped, free-form request builder with quick-call prefills and a JSON viewer for every response.
No CLI. No hand-crafted API calls. No copying tokens into curl. Just the flow you'd want.
The token never touches the browser. We store it in an encrypted, httpOnly session cookie and inject it server-side on every call.
Pick a WABA, pick a phone, click onboard. The studio provisions the agent through Meta and shows a live six-step checklist (eligibility, business info, first skill, first connector, rollout) so you always know what's next. Need to start over? A guarded Delete agent action removes the agent (and disconnects the integration when it's the last one on the WABA).
Author skills, upload knowledge, wire connectors to your existing APIs (14 SaaS templates to start from, or bring your own OpenAPI spec), or import a config JSON you already have.
Chat with the agent in the studio, run eval suites, replay scenarios against your live config, then flip the rollout switch.
The studio isn't just a form-and-list app. It's built for the messy realities of running agents in production.
Ship agent behavior without babysitting engineers for every copy tweak.
Type-safe schemas, a real API playground, and every call recorded for cURL.
Export a client's whole configuration as JSON, replicate it on the next account.
If you have another question, browse the full help center. Every page in the studio also has a Help link that opens the matching article.
A Meta Business account with the Agent Platform enabled, plus a valid access token. You paste the token on sign-in and the studio handles the rest, discovering your WABAs, phone numbers, and existing agents.
In an encrypted, httpOnly session cookie, server-side only. The browser never sees the raw token, and every Meta API call is proxied through same-origin routes that inject it at request time.
The agent itself runs on Meta's Business Agent Platform. Optional assistive features (skill drafting, doc-to-agent generation) call your configured provider: either the local Claude Agent SDK or any OpenAI-compatible endpoint (OpenAI, Ollama, LM Studio, Together, and more).
Yes. Every workspace can be exported as a versioned JSON bundle covering settings, business info, skills, knowledge, connectors, and tools. Import validates the shape, previews the counts, then applies mutations sequentially with a live progress bar.
The studio is self-hosted. Ship it as a Docker container (Dockerfile plus docker-compose.yaml included), on Apple's native container CLI, or as a one-shot Zoho Catalyst AppSail deploy. You keep control of the box, the cookie secret, and the access tokens.
Every page has a Help link in its top-right that opens the matching article in a new tab. The full help center at /help has a search box (press / to focus) that ranks matches across article titles, sections, and body text.
Use the built-in test chat, save any promising conversation as a scenario, and replay it side-by-side against the current config. Each turn is marked match, differs, or error so you can spot regressions at a glance. Combine with the Meta eval suite for structured runs.
Give them an agent that actually helps. Sign in with a Meta access token and be testing your first agent within the hour.