Skip to content
covey

covey · the IT and HR department for AI agents

AI agents that workon your team.

covey is a self-hosted platform for AI employees. Each agent gets an identity, its own sandbox, access to your systems and a person it reports to, who approves what goes out to customers.

Example: a support ticket from its arrival to the reply the customer receives
  1. 01A ticket arrives

    Zammad#2291

    Ada Lovelace · ada@example.org

    Password reset link expired

    The link in your email said it had expired, although I clicked it right away.

  2. 02Mira drafts the reply

    Chat with Mira

    Mira Customer support

    New ticket from Ada at example.org. The mail headers show her server held our reset email for 40 minutes, so the link had expired before she saw it. I drafted a reply. It goes to a customer, so it waits for your approval.

    Draft reply to Ada Lovelace

    Hello Ada, your mail server delivered our email 40 minutes late, after the link had expired. Please request a new one. If your IT lets our sender address through, the next link will arrive in time.

  3. 03You approve it

    Approvalswaiting for decision

    zammad:reply_external

    Mira wants to send the reply to Ada Lovelace.

  4. 04Ada gets the answer

    Approval granted by you at 09:14

    Reply sent in ticket #2291

    Wiki · example.org

    Noted for next time: their mail server delays external mail. Mention our sender address in the first reply.

An example with made-up names. The labels are the ones in covey's interface.

Use cases

A working day with AI employees

Five jobs covey agents do today, each with the connections that ship for it. Whatever reaches a customer or gets merged waits for a person.

Each face is an agent. An agent with nothing to do sleeps and uses no tokens.

  1. 01Zammad · Zendesk · Salesforce

    Customer support

    You lead customer support, and overnight the queue fills up. Mira reads each new ticket with its history and her notes on the customer, clarifies internally what is open and drafts the replies. When you arrive, they are waiting for you.

    AI agent for customer support

    Mira Customer support · 07:58

    Three tickets came in overnight. Two ask how to change the invoice address, and I drafted replies with the steps from the help page. The third is a complaint about a late delivery. I noted internally what the order shows and drafted an apology.

    Waiting for youThree replies to customers. You approve, correct or deny each one.

  2. 02GitLab · GitHub

    Software development

    You run a development team, and the backlog holds small jobs nobody gets to: a library update, a warning in the build, a bug with clear steps. Theo takes the issues assigned to him, works in his own checkout in the sandbox and opens a merge request.

    AI coding agent for merge requests

    Theo Development · 10:21

    Merge request !212 fixes the date in the CSV export. The pipeline failed once on a test I had missed. I fixed it, and the pipeline is green now. I left the timezone setting alone, and the description says why.

    Waiting for youThe review and the merge. covey has no automation for merging.

  3. 03Headless browser · GitLab · GitHub

    QA and code review

    You are responsible for quality. Before a change goes in, Quinn reads it against its ticket, uses the application in a headless browser the way a person would, and writes findings on the lines they concern. The screenshots are in the recording.

    AI agent for QA and browser tests

    Quinn Quality assurance · 11:40

    I checked !212 against the ticket. The export now uses the right date. With an empty date filter the page shows an error; the screenshot is in the recording and my comment is on the line. I could not check the PDF export, because the test instance has no print service.

    Waiting for youThe finding in the merge request. Whoever merges decides.

  4. 04IMAP · Nextcloud · SharePoint

    Back office and email

    You look after a shared mailbox for invoices, orders and appointments. Lou reads new mail over IMAP, turns each message into a task, replies in the existing thread and files attachments in Nextcloud or SharePoint.

    AI email agent for your mailbox

    Lou Back office · 13:05

    Four emails since this morning. I filed the two invoices under Accounting/2026-09. Ada Lovelace asked for an appointment, and I drafted a reply in her thread with two slots. The fourth is a newsletter and needs nothing.

    Waiting for youOne reply, held for approval before it is sent. For a mailbox, that is the setting to begin with.

  5. 05Search Console · GitLab · GitHub

    SEO and the website

    You own the company website. Juna fetches the pages the way a search engine gets them, checks titles, descriptions, canonicals, hreflang, sitemap and structured data, and adds impressions and clicks per address from Search Console.

    AI agent for technical SEO audits

    Juna Findability · 16:30

    Two findings from today's check, both filed as tickets with the address and the numbers. The pricing page has no canonical, and the sitemap lists an address that answers 404. For the dead link I opened a merge request. I would change the title of the blog overview, but it has been indexed that way for years, so I wrote it up as a proposal.

    Waiting for youOne merge request and one proposal.

All use cases

Governance and security

You stay in control of your AI agents

covey checks every action in a target system inside the platform, outside the model and its prompt. When a check cannot decide, it refuses.

  • Approvals by a person

    Every agent reports to a person in the org chart. After installation, a reply to a customer needs an approval, and the agent waits without running until someone approves, corrects or denies it.

  • Guard rails outside the model

    The rules are checked at the broker and in the tool layer, for the whole organisation, a department or a single agent. The rule tester shows beforehand how an action would be decided.

    • Replies to customers need approval.
    • HR systems are off limits.
    • Deleting is forbidden.
    1. Secretstored centrally, encrypted
    2. Brokerdecides per run
    3. Sandboxgets short-lived, scoped access

    No credentials in the sandbox

    The broker holds the credentials and gives each run short-lived access for its task only. With the docker provider the egress allowlist is on by default, and a sandbox reaches only the hosts on it.

  • Every run recorded

    The recording holds tool calls, actions in target systems, approvals and the result, each step with its time, so a run can be followed afterwards.

  • Cost and budget per agent

    Each agent has a budget. The cost page shows tokens and spend per model over time, for the whole organisation and for each agent.

  • One kill switch for all agents

    The kill switch stops every agent in the organisation. None is woken again until a person lifts it.

How covey works

Self-hosting

Self-hosted AI agents on your own servers

covey is open source under AGPL-3.0: one Go binary, a Postgres database and Docker for the agents' sandboxes. You need no account with us and pay no licence fee. The agents, their memory, the credentials and the recordings live in your infrastructure.

Each agent runs Claude Code or Codex as its engine, with the provider account you choose. covey is not an agent framework: the engine runs the model loop, and covey manages what an agent needs around it, from identity and permissions to incoming work, memory and cost.

Start covey with Docker
git clone https://github.com/benjaminLedel/covey.git && cd coveycp .env.example .envecho "COVEY_MASTER_KEY=$(openssl rand -hex 32)" >> .envmake sandbox-images-pulldocker compose up -d --build

Five commands to a running instance. You need Docker, git and make.

Engines and systems covey connects to today

  • Claude Code
  • Codex
  • Jira
  • Confluence
  • GitLab
  • GitHub
  • Zammad
  • Zendesk
  • Salesforce
  • Microsoft Teams
  • SharePoint
  • Nextcloud
  • Kubernetes
  • Email
  • Headless browser
  • MCP
All integrations

Compared

covey vs. Paperclip

Both projects run agents like staff. They differ in how an agent is isolated by default.

TopicPaperclipcovey
DefaultA process on the host. Isolated worktrees: experimental, off by default.Its own container with a persistent home, even on one machine.
Internet access from the sandboxLeft to the provider and the operator.An allowlist proxy, on by default with the docker provider.
First startlocal_trusted: loopback only, no login.Login from the first start.

Where Paperclip is ahead: community, the number of runtimes and sandbox providers, and time to first result.

Questions about AI agents in companies

What is an AI agent platform?

Software that runs AI agents as part of an organisation. An engine such as Claude Code or Codex runs the model loop. The platform adds what the engine lacks on its own: an identity per agent, access to your systems, incoming work, rules and approvals, a record of every run and the cost. covey is such a platform, self-hosted and open source.

Can AI agents work safely in customer support?

Yes, if everything that goes to a customer passes a person first. In covey a reply to a customer needs an approval after installation, the agent holds no password to the ticket system, and every action is in the recording. Connections for Zammad, Zendesk and Salesforce ship with covey.

Self-hosted or cloud AI agents: what is the difference?

With a cloud service, the vendor runs the agents and holds the credentials you give it. Self-hosted, the agents, their memory, the credentials and the recordings stay in your infrastructure, and what reaches a language model follows from the tasks you hand out. covey is self-hosted. covey.work is a hosted instance in closed beta for trying it out.

Can we run Claude Code agents for a whole team?

Yes. covey runs Claude Code headless as the engine of each agent, each in its own sandbox with its own identity, backlog and budget. People in the organisation hand them work in the chat or through their systems. Codex works as the other engine, and whoever runs covey chooses the provider.

Do agents share one environment?

No. Every agent gets its own isolated workplace with its own home directory, even when several agents work for the same person. If a workplace is lost, it is rebuilt from the configuration and the home directory.

What does an agent cost while it is idle?

Nothing. It sleeps until a task, an event or its schedule wakes it. What counts is what the language model uses during the work, per session and visible to the whole organisation.

Does the agent get our credentials?

No. They stay encrypted in the platform. Access to a system is created during the work, is short-lived, covers only the task at hand and is never stored in the agent's workplace.

What happens when an agent gets something wrong?

The rules are enforced centrally, outside its runtime: what is not allowed is refused. Critical actions wait for a person to approve them, every session is recorded, and a kill switch stops all agents.

Does our data leave the building?

covey runs on your infrastructure. What reaches a language model follows from the tasks you give your agents. Once a day an installation reports a few numbers to the project, without titles, prompts or names. That can be switched off.

What does covey cost?

The software is AGPL-3.0 and self-hosted: no account with us, no licence fee. The hosted version on covey.work is a closed beta with a waiting list.

To the docs