Start free
Overheard in your terminal today

Lies, damned lies and “all tests pass.”

Your agents are fast, tireless and occasionally full of it. Hydrant is where their claims meet a paper trail.

  1. Done!

    Hydrant ships with Agent Review and Human Review stages, so the handover has somewhere to wait. The workflow pack stops your agent there. It goes further only when you grant it one issue to ship, and the grant goes on the record. Either way, Done is your call.

  2. All tests pass.

    Then the output goes on the issue. The free workflow pack makes agents run your checks and get an independent review before they hand anything over.

  3. Picking up where we left off.

    It can’t remember lunch. The issue remembers the spec, the blockers, the history and the last attempt, so the next session starts from what happened.

The tracker

Less watching work.
More moving it.

Mark #42 done and watch #43 unblock itself. It’s live. Poke it. Nothing here touches a real workspace.

Launch control / WatchtowerInteractive demo
Ship before something
catches fire.
THE PLAN 3
Launch readiness1/3
HYD / 42

Fix the leak in checkout

The payment goes through. The confirmation goes missing. A spectacular way to make people pay twice.

Keep the receipt attached to the request.
Return the same result when a request is retried.
JuliusAgent

Found it. The retry path was creating a second request. Same key, same receipt. Much less water on the floor.

Go on. It’s a demo.
Try it: mark #42 done. Watch #43 unblock.

Real interactions. Imaginary fires.Nothing here touches your workspace.

  • 01

    Specs before code.

    Refined means someone settled the scope and the acceptance checks. Ready means it’s refined. Blocked is worked out for you. Your agent stops guessing which one you meant.

    How it works: Specs before code.
  • 02

    Find the bottleneck.

    Blockers, sub-issues and related links, drawn as a graph when words stop helping. Blocked work says so, loudly, and unblocks itself when the blocker ships.

    How it works: Find the bottleneck.
  • 03

    One Playbook. Every agent.

    Write the rules once: how you refine, what “done” means, what never ships on a Friday. Attach guides to issues or make them defaults, and see exactly what guidance each agent received.

    How it works: One Playbook. Every agent.
  • 04

    Receipts for everything.

    Every change carries a name: yours, or “Claude (agent; via you)”. Descriptions keep their revisions. When something moved, you can find out who, when and why.

    How it works: Receipts for everything.
  • 05

    For when “soon” stops being an answer.

    Projects with milestones, roadmap targets, cycles that roll over on their own, and Insights on what actually shipped, as opposed to what everyone remembers shipping.

    How it works: For when “soon” stops being an answer.
  • 06

    Your data can leave whenever.

    Export a whole workspace as JSON and Markdown. Trash keeps things for 30 days. Delete your account yourself, no email required.

    How it works: Your data can leave whenever.
How work flows

From “hmm” to Done.
A human at both ends.

Open issues carry a prompt button for the next step. The free workflow pack teaches Claude Code and Codex what each step means in your repository.

  1. Capture

    You or an agent

    Somebody files it. The duplicate check runs first, because the backlog doesn’t need a fourth “login is weird”.

  2. Refine

    Agent drafts, you decide

    Scope, acceptance checks, size, owner. Ready means specified, not vibes.

  3. Go

    Agent

    An agent builds it on its own branch and runs your checks. It doesn’t merge. Merging isn’t its call.

  4. Review

    Another agent, then you

    An independent reviewer tries to break it. Whatever survives lands in Human Review with the receipts attached.

  5. Done

    You

    When you say so. Not when the model feels good about itself.

Get the workflow pack. It’s free.
Bring your own agents

Bring your own robots.

Hydrant doesn’t run agents or sell you a model. Connect the ones you already pay for. They get one workspace, a name in the history and rules they can’t negotiate.

Connect over MCP
OAuth in the browser or an agent key. Tested with Claude Code, Claude and Codex.
Scoped, not trusted
One workspace, never more than the person who connected it. Revoke it whenever you like.
Named in the history
Every change says which agent made it and on whose behalf.
Taught your process
Playbook instructions and a free workflow pack give agents your rules for capture, refinement, review and release.
Connect your agent in about two minutes
agent sessionhydrant.dev/api/mcp
  1. YouPick up #42 and fix it.
  2. get_issue#42 · Fix the leak in checkout
  3. update_issue#42 · status → In Progress
  4. JuliusFound it. Retries were charging people twice. One key, one receipt now.
  5. add_comment#42 · “Fixed. 14 tests, 0 failures. Output attached.”
  6. update_issue#42 · status → Agent Review
  7. YouChecked the output. Marked #42 done.
  8. Hydrant#43 Open the floodgates is unblocked. Julius’s changes are in the history, on your behalf.
Julius (agent; via you)
Receipts

This page is issue #292.

Hydrant is built in Hydrant, by a small team and a larger number of bots. Here’s how the page you’re reading got made.

  1. Human

    Filed “Marketing site sweep”: twelve bullet points and “anything else I’m missing”.

  2. Claude

    Refined it: scope, ten acceptance checks, size XL. Marked it Ready. Blocked by onboarding and the help rewrite until both shipped.

  3. Claude

    Built it on its own branch. Ran the checks. Attached 12 screenshots to the issue.

  4. Another Claude

    Reviewed it independently. Found four blockers. All four got fixed before anyone asked a human.

  5. Human

    Liked the fine print. Sent the rest back for more edge. You’re reading round two.

Trusted by

Our logo wall is experiencing a customer shortage.

  • UNDISCLOSEDClassified
  • STEALTH MODEClassified
  • BIG NAME™Classified
  • REDACTEDClassified
  • YOUR LOGOClassified

Imaginary logos. Actual software. Hydrant is built in Hydrant, one issue at a time.

We let robots write our code.
Here’s how we keep them honest.

Read the manifesto
Questions

Before you turn the valve.

Does it run AI agents for me?

No. Bring your own agent client and connect it through MCP with a workspace-scoped credential. Hydrant exposes work and records changes; it does not supply a model or run agents in the background. Your agent provider has its own terms and data practices.

Which agents can connect?

Any MCP client that speaks Streamable HTTP, signing in through the browser with OAuth or using an agent key. We test Claude Code, Claude (claude.ai and Desktop) and Codex. Setup takes about two minutes, most of it spent in your browser.

Can I use it with a team?

Free includes 1 human per workspace; Solo Pro includes 3. The owner’s plan covers everyone in that workspace. Congratulations on the hiring. Email bots@hydrant.dev with your headcount and we'll sort out seats faster than your team can agree on a label colour.

Why do I need GitHub?

GitHub is the current sign-in provider. Hydrant does not require you to select or connect a repository to track issues. Connecting repositories and syncing issue tracking with GitHub/PR statuses is planned, but not available today. Other sign-in providers are not offered yet.

All questions
Julius’s very unnecessary pressure department

Go on. Make a scene.

Four turns. The whole page becomes your problem.

All quiet. For now.JLS—001 / HYDRANT
What could possibly go wrong.JULIUS Pressure management00 / 04 TURNSYes. The big orange one.
HYDRANT / INCIDENT #000

Well. That escalated.

Cosmetic catastrophe. Your actual work is safe.

Automatic drainage included. We’re reckless, not monsters.
Your real issues remain dry. This is strictly recreational.NO WORKSPACES WERE FLOODED.

Your agents are waiting.
Impatiently.

Free plan, no card, about two minutes to your first issue. Another two to hand it to an agent.