FAQ

Quick answers to the questions we hear most.

What is the difference between a connection, an action, and a Smart Tool?

A connection is your encrypted, authenticated link to an app (Slack, GitHub, any HTTP API) and the wiring every action uses to reach it. An action is a deterministic operation against a connected app (send the message, fetch the report), run the same way every time with no model involved. A Smart Tool is a reusable prompt that calls a model and returns text or schema-checked JSON, no app required. Full breakdown with a comparison table: Connections, Actions, and Smart Tools.

Should I build an action or a Smart Tool?

Ask whether an API can do the task exactly, or whether it takes reading and judgment. "Fetch the last 10 orders" or "create a ticket" is an action: deterministic, cheap, no model. "Summarize these orders" or "decide which look fraudulent" is a Smart Tool: it calls a model and returns words or JSON. Many workflows use both, an action to get the data and a Smart Tool to make sense of it.

Does an AI build cost $0.50?

No. That number is a refundable hold, not the price. The build settles at what the models actually used (a hold of $0.50 on a build that used six cents costs six cents) and the rest refunds automatically the moment the build finishes or fails. The ledger shows both the hold and the settlement, so you can reconcile every build to the cent.

My AI-drafted connector 404s. Do I have to start over?

No. Open the connection's detail page and use Fix with AI (it reads the recent failures itself) or Edit definition to correct the JSON by hand. Your connections pick up the fix automatically and you can retest in the Playground right away. Agents can do the same with get_connector, update_connector, and fix_connector.

Why did my agent's create_action call get denied?

Your workspace's tool policy (the default) requires a human-opened window before agents may create, rebuild, or delete tools. Open one under Settings, Agent tool changes (it closes by itself), or set the policy to Open if you want agents to build freely. Running existing tools is never affected.

What happens when my balance hits zero?

Runs and builds are denied with a clear message and a top-up link. Nothing is charged for a denied run, and the balance never goes negative. Existing connections, tools, and data are untouched.

Can I use my own Anthropic key?

Yes. Set it in workspace Settings and AI builds call the model on your account instead of the platform's.

How do I revoke an agent?

If it connected with OAuth: Settings, Connected agents, one tap. If it uses an API key: revoke the key on the Keys page. Both take effect within seconds.

Why did my run cost what it did?

Each run's detail page shows its exact cost, and the ledger links every charge to its run. Pricing rules can vary the per-run price by connector, run kind, or time of day; the Playground always shows the applicable price before you press Run.

How do I give an agent fewer tools?

Create a bundle. A bundle is a separate MCP URL exposing only the tools you assign to it, so a support bot can get three tools while your own agent keeps the full surface. See Connect your agent.

Is my data used to train AI models?

No. ToolGlue calls an AI model only when you explicitly ask for a build or a connector draft, and only with the context needed for that job. Your credentials never enter a prompt, and build transcripts are redacted before storage.

Do failed vendor calls still bill?

A run that was dispatched bills its run price even if the vendor returns an error, because the work was done. A run denied before dispatch (validation, missing scope, no balance) costs nothing.

What happens if a build fails?

The unspent part of the hold is refunded automatically. The build page shows what was tried; a rebuild with feedback usually gets there.

Can teammates join my workspace?

Yes. Invite them by email from Settings. Roles are Owner, Admin, and Member, and the last Owner can never be removed by accident.

OAuth or API key: which should my agent use?

OAuth when the client supports it (claude.ai, Claude Desktop): no key handling, short-lived tokens, one-tap revoke. API keys for CLIs, editors, scripts, and CI where pasting a key is simpler. Both carry the same scopes and behave identically once connected.

Can I connect multiple agents?

Yes, as many as you like: each OAuth grant and each API key is tracked and revocable on its own, and bundles let you give each agent a different tool surface.

Can I delete a connection?

Yes, from its detail page. Archiving it hard-deletes the stored credentials immediately.

Where do I see everything my agents have done?

The Runs page lists every call with status, cost, and duration, and each run's detail shows the recorded, redacted HTTP exchange. Money movements live on Billing; security events on Audit.