Klingbar · AI agents for small teams

A crew of AI agents that watches your business.

Klingbar runs crews of AI agents that watch your business and report in Slack. Each agent is a persona with a job, a schedule, and a channel. It runs on a machine you control, under your own engine logins, and it proposes every change for a person to approve.

#revenueTuesday, 7:02 AM
RevenueApp7:02 AM

Two annual renewals failed overnight, both on expired cards. Neither customer has been emailed yet. Details are in the thread.

AP-17 — Email both customers a card-update link

Send the renewal-failed template to the two accounts named in the thread. The drafts are attached there.

Revenue · proposed Tue · expires Fri (72h)
ApproveDeny…
An illustration: a persona's morning finding and the card it proposes. The persona waits for an approver before it sends anything.
Findings, nothing, or blocked

Silence is the normal day

Every run ends one of three ways. A finding posts a short summary — the bottom line first — with the full report in a thread. Nothing posts nothing at all, so a quiet channel means the crew looked and everything was fine. Blocked says exactly what is missing, so you fix one thing instead of guessing.

Every run ends with one
FINDINGS
Something worth your attention: a short post, the report in its thread.
NOTHING
It looked and all was well. Nothing is posted.
BLOCKED
It could not do the work, and it names why: an expired login, a missing credential, a dataset it cannot reach.
Personas

A persona is a job, written down

A persona is a prompt that says what it watches and how it decides, the tasks it runs on a schedule, the Slack channel it reports in, and a seat — the engine login, model, and effort it runs on. Give it a standby seat on a second engine, and a usage limit on one login no longer stops the work.

What a persona is made of

[personas.revenue]
display   = "Revenue"
channel   = "#revenue"
account   = "claude-a"
model     = "opus"
secondary = "codex"

[personas.revenue.tasks.morning-review]
cron    = "0 7 * * *"
timeout = "30m"

You fill in forms on klingbar.com. The box receives the result as crew.toml, versioned, and never hand-edited.

Cards

Reads are free. Changes are cards.

A persona can read anything you give it access to. When it wants to change something — send an email, push a branch, edit a record — it proposes a card. An approver decides in the Slack channel, from the Slack Home tab, or on klingbar.com. Approved, the persona makes exactly that change and reports back. Denied, or left for 72 hours, nothing happens. The rule is part of every persona's prompt; the hard limit is the access you give it.

How cards work

RevenueApp9:16 AM
AP-17 — Email both customers a card-update link
Executed · 9:16 AM · see the report

In the thread: AP-17 approved by Dana (via home)

The same card, updated in place once the change was made. The thread records who approved it and where they did — here, from the Slack Home tab.
Your box

It runs on your machine

The personas run inside one small daemon on a box you control — a server, a VM, or a spare machine — and one box can serve many crews. It dials out to klingbar.com; nothing dials in. Your engine logins and secrets stay on the box, and a secret you type on klingbar.com is sealed in your browser to the box's own key, so we cannot read it.

What stays on your box

$ curl -fsSL https://klingbar.com/install.sh \
    | sh -s -- --pair KB-XXXX-XXXX

The installer and your engine logins are usually all you type on a box. A persona that works in your code repository also needs one crew workspaces --crew <slug> there and read access to the repository, and a tool it reads through a command-line login, such as gcloud or the aws CLI, is signed in on the box. Configuration, prompts, secrets, and upgrades all come from the web.

klingbar.com

Everything it did, on the record

klingbar.com keeps the crew's record: every run with its verdict, transcript, and what it spent as the engine reported it; every card and who decided it; open questions; learnings; weekly digests; the schedule; and the box's own health. What the box did not report reads "not reported" — never a green nobody measured.

Tour the platform

Example run record
persona
Revenue · morning-review
verdict
FINDINGS · 1 card proposed
seat
claude-a · opus · high
turns
14
cost
$0.61, the engine's own estimate
Learnings

It learns your business, with your say-so

When a run finds something worth remembering — a vendor that always invoices late, a metric that misreads on Mondays — it proposes a learning. Once an approver accepts it, every future run of that persona carries it. Standing rules do the same for the whole crew, and each persona's weekly digest says what it did.

RevenueApp7:05 AM
L-8 — proposed learning

Annual renewals retry once, three days after the first failure. Check again on day four before emailing anyone.

ApproveReject
An illustration of a proposed learning. Approved, it joins every future prompt for this persona.
Focus groups

A marketing persona can ask a room first

A crew's marketing persona can put its creative in front of a synthetic cohort drawn from our public library of 197 panelists, run three or more rounds on your box, and propose the winner as a real experiment where your app is wired for one. During early access we set that persona up with you. Synthetic reactions are hypotheses; the experiment is the evidence.

Start with a small crew

Most crews start with a chief of staff, whose brief always posts so you can see the crew is alive, plus the personas the guided setup suggests for your business — revenue, errors, support — and grow from there. Klingbar is free during early access: you bring the machine and your own engine logins. What that means.

Request early access