Run many AI agents, and know what each one did.
An AI workforce is less one clever agent than dozens of scheduled runs a week, on several engines, for more than one part of the business. Klingbar runs crews of AI agents that watch your business and report in Slack, and klingbar.com keeps the record of every run: when it ran, on which seat, how it ended, and what it spent.
One box, many crews, many personas
-
Your box
One daemon
- Serves every crew on the machine
- One cap on runs across the whole box
- One engine login never runs two jobs at once, across crews
- serves
-
Crews
One per business or team
- Its own channels and approvers in your Slack workspace
- Its own configuration and standing rules
- Its own cap on runs at once
- each runs
-
Personas
One per job
- Tasks on cron schedules, or on demand
- A seat, and a standby on another engine
- A channel it reports in
A calendar you can read
Every task carries a cron schedule, or none for on-demand work, and each crew's schedules page lists every task with when it is next due and when it last ran. Runs queue on the box and start as the caps allow: a limit per crew, one for the whole box, and one run at a time per engine login, across crews.
If the box was down when a run was due, it catches up once on return, within a window the crew sets. From the roster you can pause a persona, resume it, or switch its schedules off, one persona at a time.
- morning-review
- Revenue · daily at 7:00 · next Tue 7:00
- inbox-sweep
- Support · every two hours on weekdays · next 10:00
- error-review
- Logs · daily at 6:00 · last run found nothing
- ideate
- Marketing · Mondays · persona paused
What each run spent, as the engine reported it
Every run record carries what the run spent, as its engine reported it: tokens, turns, and the engine's own cost estimate, summed over every attempt — including runs that timed out or were stopped, because those spent real money too. klingbar.com computes no price from tokens, and an engine that reported no figure reads "not reported", never zero.
klingbar.com records spend; it does not budget it or block it. What runs cost is what your engine subscriptions charge, and the controls that bound a run are its schedule, its timeout, and a turn ceiling.
Est. spend, 7 days: $38.40 across 41 of 44 runs that reported it
- inbox-sweep
- Findings · 9 turns · $0.21, the engine's estimate
- error-review
- Timed out at 45m · 31 turns · $0.88
- upgrade-posts
- Nothing · cost not reported
A turn ceiling for runs that wander
A crew, or a single task, can set max_turns: the most turns a run may take. It applies on Claude Code and Grok, the two engines that take the flag. A run that reaches it ends with its own verdict, CEILING, and is neither retried nor restarted in a fresh session, so one runaway run costs one run's worth of turns.
Every task also has a timeout. A run that passes it is stopped and recorded as timed out, with what it had spent so far.
[crew] max_turns = 40 [personas.logs.tasks.error-review] cron = "0 6 * * *" timeout = "45m" max_turns = 25
Set on klingbar.com, on the crew or on a task. Claude Code and Grok only.
A usage limit pauses one login, not the crew
A usage limit is a fact about a login, not about a crew, so that is where it applies: the limit pauses only that login, the job keeps its session, and it resumes when the window reopens. Every crew whose personas use that login sees the same pause.
Give a persona a standby seat on a different engine and its run moves there instead of waiting; the primary is preferred again at the next run. Two accounts on the same engine login are refused as a standby, because a pause on one is a pause on both.
- claude-a
- Paused until 2:00 PM · usage limit
- support
- Running on its standby seat,
codex - revenue
- No standby · queued, resumes on
claude-aat 2:00 PM
Health that says "not reported" instead of guessing
Each box page shows its host, daemon version, installed release, the engines and logins it found, the crews it serves, and its health checks (crew doctor). A box that stops checking in is marked stale after five minutes, and you get an email; a box that cannot prove its own home directory refuses to write and says why.
Anything the box did not report reads "not reported" — never a green nobody measured, and never a zero the box did not send.
- crews
- 3 served · last check-in 20 seconds ago
- engines
- claude, codex, grok · gemini not reported
- logins
claude-apaused until 2:00 PM- health checks
- 1 warning: the app has not joined
#support
Moving a crew to another box
When a crew outgrows its machine, move it from the crew's danger settings. The current box re-seals every secret to the new box's key and sends the crew's jobs, threads, cards, and state with them. Logs and workspaces stay behind: klingbar.com already has the transcripts, and on the new box, crew workspaces rebuilds each workspace from its remote. Nothing is deleted first — the old box keeps the crew until the new one has it, and only then retires its copy.
How this differs from orchestration
An orchestrator answers a runtime question: given this task, which model or tool runs next, with what input, and what happens when it fails. It is a scheduler with a graph in it, and if that is the problem you have, an orchestrator is the right tool.
Managing a set of agents answers a different question, and it is the one that arrives second: who is allowed to do this at all, whose approval it needs, what it spent, and where the record of that decision lives next month. None of those are properties of a single run — they persist between runs and outlive any single persona.
The two stack rather than compete. Klingbar does not replace the engines underneath its personas, and a task can drive whatever tools you give it. For the full distinction between a control plane and a workflow engine, read the guide to the agent control plane, or see what runs on your box and what crosses the wire.
Start with one crew, then add the next
Klingbar is free during early access. One box can serve every crew you add; you bring the machine and your own engine logins. What that means.
Request early access