FAQ
Questions, answered
What people ask before and after they build their first agent. Can't find yours? Write to us.
A platform for AI agents that do real work. They use your apps, read your files, run on a schedule and talk to people on Slack, WhatsApp, Telegram, Chrome and a web page of their own. Every step is traced, and risky actions can wait for a person.
No. You describe what the agent should do and Builder sets it up: instructions, apps, approvals, skills and schedules. It tries the agent on a few questions as a dry run, where actions don't really happen, and you publish when you're happy.
The agent that builds and fixes your agents. It asks when something is unclear, with the answer it recommends, drafts the agent part by part and tries it before you publish. It never publishes on its own. Builder is on every plan, Free included.
Sign up on the Free plan, add a key for one model provider, and ask Builder for your first agent.
Sign up freeYes. Invite people by email as owners, admins or members; an invite lasts 7 days. For now each person belongs to one workspace.
Upload files to a knowledge base: PDFs, Word, Excel and PowerPoint files, markdown, CSV, JSON and more, up to 10 MB each. The agent searches them by their words and walks the folders, reading the files a question needs. Scanned PDFs aren't read yet.
A skill teaches an agent how your team does a task: instructions, and sometimes scripts, in a folder. Add one from the open catalog at skills.sh, upload your own, write one in the editor, copy one from a public GitHub repo or ask Builder to draft it. Catalog skills show what five security auditors found, and the agent's owner approves every skill.
Yes. A task runs once at a set time, or daily, weekly, monthly or on any cron schedule, in your time zone. Each run starts a new chat you can open later. Builder can set one up for you.
Yes, any remote MCP server over HTTPS (Streamable HTTP or SSE), with no auth, a bearer token, an API key header or OAuth. Local stdio servers aren't supported.
Every agent can search the web and read pages, on by default. With the Chrome extension it can also work in your browser: read the page, click, type and move between tabs, only on sites you allow, and hand back to you when it needs you.
Each conversation gets its own Linux computer where the agent can run code and commands. It's kept between messages, stops after 15 minutes idle and is deleted 30 days after it was last used. Secrets you add become environment variables, hidden in the output, and only your team's own chats and schedules get them.
Edits go into a draft; people keep talking to the published version until you publish. Each version keeps its own model, skills and sandbox settings. Before you publish, evals run on the draft and show anything that got worse.
A run takes up to 25 steps. A message holds up to 32 KB of text and up to 10 files of 10 MB each. A conversation answers one message at a time; the next one waits its turn and is never dropped.
Bring your own keys for Anthropic, OpenAI, Google Gemini, DeepSeek, Kimi or OpenRouter. Each message can go to a fast or a slow model, so simple questions stay quick and cheap.
Calls made with your own provider keys are billed by your provider, never by us. Builder, the agent that builds agents, runs on our keys on every plan.
Give an agent a fast model beside its main one. For each new message a small router decides which should answer, and either model can hand the work to the other mid-run. The router is free, and its choice shows in the trace.
Read how it worksKeys are encrypted with AES-256-GCM and never appear in logs or traces.
Not yet. Your agents run on your own provider keys; only Builder runs on ours.
Over 1,000, through Composio: Gmail, Slack, GitHub, HubSpot, Notion, Google Sheets and many more. Apps can start an agent too: a new email, a new row or a webhook can trigger a run.
You connect each app yourself, or each customer connects their own account. Actions that write or send can wait for a person to approve them, and every call is in the trace.
Yes. Set an app to use the customer's account, and each person connects their own on the agent's web app. The agent never falls back to yours.
Slack, WhatsApp, Telegram, the Chrome extension and a web app that comes with every agent. More channels are on the way.
Every agent gets a page of its own that you can share with customers. You decide who gets in: sign-in with Google or an email code, limited to your domains if you like, and questions people answer before they start. Leave sign-in off and people come in as guests.
You link a WhatsApp number the way WhatsApp Web does, and the agent answers one-to-one chats on it. It's an unofficial connection, so WhatsApp can ban the number; use one kept for the agent.
You create a Slack app in your workspace from a manifest we give you, and the agent answers there. Questions and approvals arrive as messages; reply in words, or YES or NO.
Write eval cases with checks, including an LLM judge, and compare every change against a baseline. Evals run on the draft before you publish, so you see anything that got worse first.
Words the answer must or mustn't contain, a pattern or valid JSON, tools it must or mustn't call and their order, step and cost budgets, and an LLM judge that grades against your rubric. Actions that write or send answer from fixtures, so nothing really happens.
Every model call and tool call in order, with inputs, outputs, tokens, cost and timing, which model the router picked and why, and the ratings people gave the answer.
List the actions an agent must ask about. When it wants to take one, the run pauses until the person in the chat says yes or no: an Approve or Deny card in the app, or a YES or NO reply on Slack, WhatsApp or Telegram. A scheduled run asks the agent's publisher.
It files a report: a missing tool, unclear instructions or a failure. You read reports on the agent's Reports tab, and Fix with Builder drafts a change for you to review. Builder never acts on a report on its own.
We're certified to ISO/IEC 27001 and ISO/IEC 42001 and comply with India's DPDP Act. Our SOC 2 audit is in progress.
Visit the Trust CenterApp credentials and provider keys are encrypted with AES-256-GCM. Sign-in is with Google, GitHub or an email link, so there are no passwords to steal. Each conversation runs in its own sandbox, and we tell you about a breach within 48 hours.
How we keep your data safeNo. We don't use your content to train or fine-tune AI models, ours or anyone else's. Calls on your own keys follow your agreement with that provider.
Read the Privacy PolicyOur servers run in Germany and stored files are kept in the US. A conversation also goes to the model provider its agent uses and the apps it calls.
See our subprocessorsWrite to hello@definable.ai and we erase your account or your workspace's data within 30 days of the request.
Credits pay for what runs on our side: Builder, and the sandbox computer by the second. Model calls on your own keys never use credits.
New runs that need credits won't start until you top up or your plan renews. Nothing already finished is lost, and you can buy more credits at any time.
Plans are billed monthly in US dollars.
Sign up on the Free plan. Its credits are added to your workspace the first time you use them.
Still have a question?
Write to us and someone from the team will answer.