Human-in-the-loop for AI agents

You delegate. Your team decides. Your AI agent acts, on a verified yes.

When your AI agent needs a fact only your team knows, or a sign-off, Anchord works out who is responsible, from who has answered and decided before, and asks them in Slack, inside the workflow you already run. Your AI agent finishes the job, and you keep a record of who decided.

Set up your AI agent Talk to us

Works with Claude Code, Cursor, Codex, Claude desktop and GitHub Copilot in VS Code over MCP, or any bot or automation through the REST API. Nothing else in your stack changes.

Why it matters

Your AI agent stops where your team's knowledge starts.

Built for teams that already run an AI agent or automation, but keep it on a short leash because they can't fully trust it yet. Anchord connects it to the knowledge and judgment only your people have, so your team can trust it with more.

  • It doesn't know who to ask

    Your AI agent hits something only a teammate knows, so it hands the whole job back or makes something up. With Anchord it finds the person responsible, asks them, and carries on with their answer.

  • You can't let it act

    Refunds, deploys and purchases stay manual because nobody can check what the AI agent is about to do. With Anchord it acts on a signed yes for exactly that action and amount.

  • Nobody knows who decided

    Answers live in DMs and approvals in someone's memory. With Anchord every question, answer and yes is recorded with who and when.

Where it fits

Add a human checkpoint to the AI workflows you already run.

Anchord isn't another tool for your team to work in. It slots into the flow you have at the moment your AI agent needs a person, and your helpdesk, code and tools stay as they are.

  1. Work comes inA ticket, a task or a request, wherever it lands today.
  2. Your AI agent works on itIn the tool it already runs in.
  3. AnchordIt needs a fact or a sign-off
    1. Find who's responsible
    2. Send it to them
    3. They respond
    4. Back into the workflow
  4. Your AI agent finishesIt acts, cites who answered, and the record is kept.

  • Support

    Your support bot in Zendesk or your own helpdesk

    Today
    Refunds over the limit go to a person, who reads the whole ticket and starts over.
    With Anchord
    The bot asks the account owner whether the order qualifies, gets approval for the amount, and issues the refund.

  • Engineering

    Claude Code, Cursor, Codex or GitHub Copilot

    Today
    The AI agent guesses at something only the team knows, or stops and waits for you.
    With Anchord
    It asks whoever owns it, like whether staging is back up, and gets sign-off before it deploys.

  • Operations and finance

    A Zapier step or an internal workflow

    Today
    A purchase over the limit sits in someone's inbox until they notice it.
    With Anchord
    The budget owner approves it in Slack, and the workflow places the order.

Before you start

What you need, what to expect, and what’s included.

An AI agent, in Anchord, is any software doing work for you that can call Anchord. That covers apps like Claude desktop, Claude Code, Cursor, Codex or GitHub Copilot in VS Code, connected over MCP, and your own bot or automation, like a Zendesk trigger or a Zapier step, through our API. It asks your team through Anchord and waits for the answer or the yes. Your team never talks to it directly; they only see Anchord in Slack. Work that starts with a person instead? That's Anchord for teams.

You need

  • An AI agent doing the work: one that uses MCP, like Claude Code, Cursor, Codex, Claude desktop or GitHub Copilot in VS Code, or any software that can call a REST API, like a Zendesk trigger or a Zapier step.
  • Slack, where your team answers. Someone who can add apps connects it; your teammates install nothing.
  • An Anchord API key, created during setup and pasted into your AI agent once.

What to expect

  • Anchord only sees what your AI agent sends it. It doesn't watch your other tools, and it doesn't act itself: your AI agent acts, on a signed yes.
  • You confirm who is asked each time, unless one of your rules names them. Early on, Anchord may not know who to suggest, so you name someone and it learns from who answers.
  • Answers come from people, so they take time. Questions go out in working hours, weekdays 8:00 to 19:00 by default, and your AI agent checks back. Approvals go out any time.

In this version

  • AI agents connecting over MCP with an API key, or through the REST API.
  • Asking a person, with you confirming who, and approvals with a signed, single-use yes.
  • Your team answering and approving in Slack.

Not yet

  • Apps that only connect through an OAuth sign-in, like ChatGPT connectors.
  • Answering in Microsoft Teams or email.
  • Unattended bots choosing who to ask: they name the person, or rely on your approval rules.

How it works

Four steps, inside the workflow you already run.

Say a customer asks your AI support agent for a refund. It finds out from the right person whether the order qualifies, then asks for approval with that answer attached, and acts only on a yes.

  1. Find who's responsible

    Your AI agent needs a fact, a clarification or a sign-off. Anchord works out who should answer or approve, from who has answered and decided before, and you confirm before anything is sent. Your rules can name the approver instead.

  2. Send it to them

    The question or approval goes to that person in Slack, with the context and the AI agent's best guess.

  3. They respond

    They answer, approve or decline, or say it isn't theirs, often in one tap. Answers are screened before the AI agent reads them.

  4. Back into the workflow

    Anchord returns the answer, or a signed yes for exactly that action and amount, to your AI agent, which carries on where it stopped. Everything is recorded.

Actions

Anything your AI agent shouldn't do alone.

Anchord knows these kinds of action, so a rule for refunds also covers a credit note. Anything else can still ask for approval.

  • Refunds and credits
  • Discounts and pricing exceptions
  • Purchases
  • Deploys
  • Access requests
  • Deletions
  • Messages to customers

Trust

Your team stays in charge.

  • You confirm who's asked

    Nothing reaches a teammate until you've confirmed who, unless one of your own rules names them.

  • Paced, not noisy

    A person gets only a few questions at a time and none outside their working hours. Approvals go out any time, unless you turn that off.

  • Rules you can see

    Set who approves what, or accept the rules Anchord suggests from who keeps deciding.

  • Answers are checked

    Every answer is screened before the AI agent reads it. Anything suspicious is held for you to release or discard.

  • A full record

    Every question, answer and decision is logged with who and when, and approvals export as CSV.

For your IT team: what Anchord reads and writes, and the permissions it asks for

Start with one workflow.

Pick one workflow, like refunds over a limit. Connect it in an afternoon: Slack, an API key and one line in your AI agent. Then measure how many more cases your AI agent finishes on its own. Anchord shows how often people answer and how fast, how often your AI agent's guess was right, and whether it asked the right person first time.

Set up your AI agent Or talk to us