Skip to content

Custom Actions

An agent that acts,not just answers.

Your docs explain how things work. Custom Actions let the agent reach into your own systems to answer the questions only your backend knows, and do real work: look up an order, check a plan, extend a trial. Point it at an HTTPS API you already have, no shim to build.

Live against your API in minutes

No shim to write, no server to deploy. Point at an endpoint you already run and let the agent reach it when it matters.

1

Point at your API

Enter an HTTPS endpoint you already have and pick the method. Direct Mode talks to it straight, no shim to deploy.

2

Attach one credential

Add your own API key or token once. ReleaseDock stores it encrypted and sends it on every call, never to anyone else.

3

Describe what it does

Write a plain-language description and map the inputs. The agent reads it to decide when the action is the right move.

4

Test, then publish

Preview the exact request, run it from the test panel, and publish. Drafts stay private until you flip them live.

No rules to wire

The model decides when to call

You don't write triggers or phrase lists. The agent weighs each customer question against the description you wrote for every action and decides on its own whether to reach for one. It works across paraphrases, typos, and other languages, because it fires on meaning, not on a matched string.

Verified identity

It only ever acts as the person asking

When an action looks up or changes data for a specific user, ReleaseDock resolves who is asking on the server and sends it as a value the model can never set or override. Your endpoint trusts the call because your credential authenticated it, then authorizes that identity against your own data.

  • Identity resolved server-side from the signed-in user
  • The builder blocks any chat value from deciding whose data returns
  • Built-in guardrail against IDOR-style data exposure

Full HTTP support

Read a record or change one

Custom Actions speak the whole verb set. The agent can fetch live data with a GET or do real work with a POST, PUT, PATCH, or DELETE, all over HTTPS, all scoped to the single endpoint you named. Add an idempotency key and a write stays safe to retry.

  • GET, POST, PUT, PATCH, and DELETE
  • HTTPS only, screened to reach just your endpoint
  • Idempotency keys keep writes safe to retry

Everything else you get

Direct Mode

Point straight at an endpoint you already run. No middle layer, no second server to maintain.

Start from a recipe

Pick a common job, order status, plan lookup, cancel or extend, and edit a prefilled draft instead of starting blank.

Visual input mapping

Map each field from the conversation, the verified identity, or a fixed value with a dropdown. No payload hand-coding.

Pick what it reads

Choose exactly which response fields the agent may use, so it answers from the data you intend and nothing more.

Draft and live versions

Edit safely against a draft, see an unpublished-changes badge, and publish when ready. Disable or soft-delete anytime.

Execution log

Every call is recorded with its request, response, and outcome, so you can see what the agent did and when.

Questions, answered

The practical details, before you sign up.

faq.md
What is a Custom Action?

It's a tool you give your AI agent that calls your own HTTPS API mid-conversation. Your knowledge base explains how your product works; a Custom Action answers the questions only your backend knows, like order status, current plan, or usage, and can take actions such as extending a trial or resetting a key.

Do I need to build a new endpoint?

No. Direct Mode points straight at an HTTPS endpoint you already run, with no shim to write and no server to deploy. If you genuinely need a transformation layer, the builder can generate a ready-to-deploy scaffold, but that path is optional and most setups never need it.

How does the agent know when to call an action?

It decides on its own. You write a plain-language description of what each action does, and the model weighs every customer question against those descriptions. There are no triggers or keyword lists to maintain, so it works across paraphrases, typos, and other languages.

How many secrets do I have to manage?

Exactly one: your own API credential. ReleaseDock stores it encrypted and sends it on every call. There is no second signing secret to rotate, and the end-user identity is never a token you have to mint.

How do you keep one customer from seeing another's data?

ReleaseDock resolves who is asking on the server, from the signed-in widget user or the conversation contact, and sends it as a value the model can never set. The builder won't let a value from the chat decide whose record is returned, which closes the most common data-exposure hole. Your endpoint then authorizes that identity against your own data.

Which HTTP methods are supported?

GET, POST, PUT, PATCH, and DELETE, all over HTTPS. The agent can read live data or change it. Add an idempotency key and a write stays safe to retry.

Can I test an action before going live?

Yes. Preview the exact request the agent will send, copy it as cURL, and run it from the test panel. Actions stay as private drafts until you publish, and you can disable or soft-delete one at any time. Every live call is recorded in an execution log.

markdown7 entries

Give your agent the keys to your backend.

Free for 7 days. Point at your API. No code to deploy.

Free for 7 days · No credit card · Point at your API