Skip to content

deliverd-dev/deliverd

v1.3.0MIT

Give any AI agent a way to ask a person — for approval, a decision, an answer or a review.

approval

Stop and ask a person before doing something consequential. Use it for anything irreversible, anything that spends money, anything that reaches people outside the conversation, deletes data, changes production, grants access, or goes beyond what you were asked to do. It covers when to ask, how to write a request a busy person can decide from a phone, how to wait, and how to act on the answer (including an approver's corrections). It works with the Deliverd MCP server, its SDKs or CLI, or in plain conversation when none is set up.

artifact

Build a beautiful, self-contained HTML artifact — report, dashboard, diagram, executive summary, plan or prototype — then offer to publish it to a shareable URL. Use whenever someone asks for a report, a dashboard, a summary, a write-up, a plan, a one-pager, a diagram, a mockup or a working prototype, or asks to "make this shareable", "turn this into a page", or "publish this".

checkpoint

Put the approval step on the path to a consequential action, in the system an AI agent runs in, rather than trusting the agent to ask. Use it when building or wiring up an agent, an MCP server, a tool, a workflow or a CI pipeline that can do something consequential (pay, deploy, delete, merge, email outside the organisation, change access); when someone asks to make approvals compulsory, put a gateway in front of tools, add a human step to a pipeline, or stop agents going around a check; and when reviewing an agent setup for where a check could be skipped. It covers choosing the checkpoint, the hosted MCP gateway, a gate in the code that acts, checking the signed decision token where the action happens, the GitHub Action, Amazon Bedrock AgentCore Gateway, and what each one does not cover. It works with Deliverd's MCP gateway, SDKs and public integrations.

compliance

Do AI work under an organisation's AI governance, and leave a record an auditor can check. Use it when the work touches customer or personal data, production systems, external communication or decisions about people; when an organisation runs Deliverd Compliance Mode; when a build step must prove an evaluation or red-team run before a release; when a model, prompt or guardrails change; and when an AI system does something harmful, wrong or unexpected (including DORA incident reports). It covers the governance profile, controls, evidence and scored eval results, the release fingerprint, the governance gate, incidents, framework mappings and the evidence pack, and what only a person may decide. It works with the Deliverd MCP server, its SDKs or CLI, or in plain conversation when none is set up.

ethics

Notice when a task could harm or wrong a person, and put that concern in front of a human before acting, never around them. Use it whenever an agent's work touches someone's job, money, housing, health, personal data, children, or reputation; speaks in someone else's name; or cannot be undone. Also use it when an organisation's ethics rules flag or refuse a request, and when helping someone write those rules. It works with the approval skill and Deliverd, or in plain conversation.

ethics-rules

Help a person write, test and roll out their organisation's ethics rules, the rules that put a concern about harm to a person in front of whoever approves an AI agent's work, ask more people to agree, or refuse it. Use it when someone asks to set up ethics rules, ethical guardrails or responsible-AI checks for agents; to adapt the starter rules; to cover hiring, credit, debt, vulnerable customers, personal data or telling people that AI is involved; to work out why a rule flags too much or too little; or to connect the rules to their AI register and policies. The agent drafts rules; a person adds them and switches them on. It works with Deliverd, or in plain conversation.

policy
prompts

Use the organisation's approved prompts instead of your own words when one fits the job, and turn words that worked into a prompt a person can approve. Use it when asked for a kind of work an organisation does again and again (summarising a contract, drafting a customer reply, reviewing a supplier, writing release notes), when someone mentions "the approved prompt", "the prompt library" or "the one we agreed", and when a prompt you wrote has worked well enough that others should reuse it. It covers finding a prompt, taking the version in force filled in with your inputs, checking the result against its limitations, and drafting a new prompt or a new version for a person to approve. It works with the Deliverd MCP server, its SDKs or CLI, or in plain conversation when none is set up.

screen

Check text for personal data, secrets and confidential terms before it goes to an AI model or leaves the machine, and redact what should not go. Use it when building anything that sends prompts to a model (wire the check into the code), when passing logs, tickets, customer records, .env files or documents on to another model, agent, API or tool, and when writing prompts, eval sets or fixtures that will be committed or shared. Runs locally with Deliverd Screen (free, MIT, no account).

tasks

Keep long AI work aimed at its goal. Use it at the start of anything substantial (more than a few steps, or likely to involve debugging or research), when you resume work after a restart or a new session, and when a side problem starts to take over the conversation. It covers stating an objective, keeping a short working state instead of relying on the history, branching a problem into a side task and bringing its result back, recording findings a person can check, writing a status report from the task's record and keeping it with the task, relying on and proposing the organisation's knowledge, running its playbooks, and leaving decisions to a person. It works with the Deliverd MCP server, its SDKs or CLI, or in plain conversation when none is set up.