Deliverd

Overview

Connect an agent to Deliverd: approvals, reviews, requests, gates and tasks over MCP, the REST API, the TypeScript and Python SDKs, and the CLI.

You build the action. Deliverd handles the human. An agent that needs a person — an approval, a review, an answer, a decision — asks through Deliverd and carries on when they respond. It can also publish what it made to a secure address behind your organisation's sign-in. These pages are for the people who connect agents; the guide is for the people who answer them and run Deliverd.

The whole integration

import { deliverd } from "@deliverd/sdk";

const decision = await deliverd.approve({
  title: "Deploy to production"
});

if (decision.approved) await deploy();

What is not in that sample: the approval page, the email and the reminders, sign-in, who is allowed to decide, the audit trail, and the screen a person decides on. None of it is yours to build.

What you can build

Ask before acting
An agent about to spend money, deploy or contact a customer asks named people to approve it, and acts on what they approved. See Approvals.
Ask the policy first
An agent declares a registered action, and rules your organisation wrote allow it, refuse it or send it to a person. See The gate.
Ask for judgement or facts
Put a draft in front of reviewers, or ask people typed questions and get the answers back as data.
Keep the record
Tie a job's requests together in a flow and export who decided what, and when, as an evidence pack.
Publish
Put a report, dashboard or prototype at a persistent address, with versions, comments and live data. See Publishing reports.

Choose how to connect

UseWhenCredential
MCP serverThe agent runs in Claude Code, Codex, Cursor, ChatGPT or another MCP client, and you write no code.OAuth, approved in a browser
REST APIAny language or platform. Everything the others do is here.An API key or OAuth token
SDKsYou write the agent, in TypeScript or Python. The waiting, retries and polling are done for you.An API key
CLIA terminal, a shell script or a CI job: block a deploy until a person approves, push actions and policy, or publish on every build.deliverd login, or a token in CI

They are one service underneath: the same scopes, the same refusals, the same audit trail. Mixing them is normal — an agent that asks through MCP can be governed by a policy pushed from the CLI.

Install

TypeScript SDK

npm install @deliverd/sdk

Python SDK

pip install deliverd

Every AI tool on this machine, via the CLI

npx deliverd setup
claude mcp add --transport http deliverd https://deliverd.dev/api/mcp

The approval skill

npx skills add deliverd-dev/deliverd-skills --skill approval

In these docs

AuthenticationAPI keys, agent keys, OAuth and scopes.SDKsAsk in one call, wait, act on the answer.Approvals, reviews and requestsThe REST API for asking a person.The gateAsk your organisation's policy before acting.Agent policyGate rules, agent permissions and standing rules.Ethics concernsWhat a flagged or refused request means for an agent.Webhooks and callbacksBe told instead of polling.In your own productA hosted or embedded decision, and an approver's inbox.MCP serverConnect each client, and every tool.CLIAsk from a build step, push policy, set up, publish.SkillsOpen-source skills that teach an agent when to ask.Deliverd ScreenOpen-source pre-send check for personal data in prompts.Publishing reportsVersions, sharing, live data, comments.Errors and limitsError codes, plan limits and rate limits.