A self-hosted builder of independent AI agents. Budgets, approvals and memory come built in.
An agent is a single job in plain sections: trigger, inputs, task, output, actions and limits.
Agents never call each other. One publishes a record, and another wakes on it.
8 kinds of gates sit between agents: condition, schedule, filter, switch, delay, join, limit and wait.
Build one agent with a mission of its own, or set up a whole team of them. The parts are the same.
One agent, or a few
Describe one job in plain sections, give it a trigger, limits and a say in what it may do. Let it work alone, or link a few: each leaves a record, and the next one wakes on it.
A company
A wizard lays out a team: 10 roles to start, and up to 32 in all. Each employee is a project with a main agent, its own budget and instructions, and you decide who may ask whom.
You stay the owner: important actions wait for your approval. A company that runs itself is a goal, not a fact yet. The main agents answer from the model and do not run code; code-writing runs separately and you start it by hand.
What is ready and what is notScreens and recordings from a demo instance.
The canvas. Agents drawn in a project and linked by control edges, with a condition gate between them.
Real product screens. Demo data only.
Anything important starts as a draft. You read it, change it if you need to, and approve it.
Every action is read, draft or act. Set what the connector allows and what the agent asks for. The lower level decides.
Budgets sit on every run, on every agent per day and on every project per day. Set a cap and start a simulated run that wants a lot.
This run wants $1.60
Spent $0.00 / $1.00
A simulation with example amounts. Template defaults are one dollar per run and five per day.
What keeps the bill small, and where you can see each part at work.
How it saves
Every input is hashed. An input the agent has already handled is skipped, so the model is not asked about it again.
Where you see it
The run page shows items deduped, and the overview shows items skipped.
How it saves
An identical request at temperature 0 gets the stored answer for 24 hours instead of a new call. Templates start with this switched off.
Where you see it
Cache hits on the run page, and the saved amount on the overview.
How it saves
The unchanging beginning of a prompt is cached by the provider, so repeated runs pay less for the same prefix.
Where you see it
The overview shows what the savers kept.
How it saves
In auto mode the router picks the cheapest enabled model that has the tools, vision, context and quality the task needs, and moves up only when a model fails.
Where you see it
The choice and the alternatives are written to the run log.
How it saves
For each kind of task LOOPmKit remembers which models succeeded and starts the next run on one with a proven record: 14 days, at least 5 attempts and an 80 percent success rate by default.
Where you see it
You can ban or pin a model by hand.
How it saves
Budgets sit on every run, on every agent per day and on every project per day. A run over its cap is skipped with a budget error, and parallel calls reserve their share first.
Where you see it
The error on the run page, and spend against the daily budget on the overview.
How it saves
Each run records tokens in and out, cost, provider and model.
Where you see it
The run page, and costs by agent and by model on the overview.
How it saves
Ollama, LM Studio or vLLM answer on your hardware, with no API key.
Where you see it
A local provider is set up like any hosted one.
Cost is only as exact as the prices set for the provider.
23 templates. Start from one task agent, or set up a team of role agents in one step.
This is how a job is described: plain sections for trigger, inputs, task, output, actions and limits. An example, with simplified field names.
agent: morning-brief
trigger:
cron: "0 8 * * 1-5"
timezone: Europe/Belgrade
inputs:
- collection: news-notes
task:
model_tier: cheap
tools: [search, fetch]
prompt: "Write a five-line digest."
output:
schema: digest.json
actions:
- send_message: draft
limits:
per_run_usd: 1
per_day_usd: 5
timeout_s: 600
memory: offThe action is a draft, so it waits in the approval queue for a person.
The result is validated against a JSON Schema before it is stored.
A run over its budget is skipped with a budget error.
Short answers first. Open a block for the details.
LOOPmKit is self-hosted: a Docker Compose stack, so agents, records and credentials stay where you run it.
Product telemetry is off by default. Model calls go to the provider you configure, or stay local if you use a local model.
Five kinds of buyer, in their own words, and the part of LOOPmKit that answers each pain.
Every task lands on me.
One secretary role takes the request and routes it to specialist employees, and you only approve. Live acceptance of this chain is still partial.
The AI bill is a surprise every month.
Budgets on every run, agent and project per day. Savers skip repeated work, the router takes the cheapest model that fits, and the overview shows the spend.
Receipts and bookkeeping pile up.
A ledger with sums done by SQL and receipts read from their fiscal QR code. The presets are for Serbia.
Enquiries wait too long.
An account manager answers from the public layer of knowledge and keeps a card for each enquiry. Support answers from what may be said.
The automation ran twice, or quietly did nothing.
Runs are durable. A failed run resumes from its checkpoint without repeating side effects, and an agent that fails three times in a row pauses itself.
I cannot tell what the agent actually did.
A run page with every step, a journal of tool calls, eleven alert rules and a feed of errors grouped by signature. The error feed is built and not switched on yet.
Mail and digests eat the morning.
Base templates for sorting mail, important mail to a chat and a morning brief. Live checks with real accounts are partial.
Content takes a whole team.
A chain from text to voice to a subtitled video, checked on a running stack up to a dry run.
I am afraid something goes out without my say.
Nothing is published without approval. One approval covers one exact publication, duplicates are skipped, and quiet hours and limits apply. No real post has gone out through the product yet.
I want specialist drafts and reviews, not a generic chat.
Roles with stack skills for Go, PHP, Vue and databases. Code runs in isolated jobs on a git branch with no push. It was tried with a stand-in for the vendor tool, not a real subscription.
We do not want to depend on one model vendor.
Twenty-five provider presets, local models among them, a cascade that learns, and a fallback.
The data has to stay with us.
A Docker stack on your machines, PostgreSQL, secrets in an encrypted vault, and product telemetry off by default.
An outward-facing agent must not read internal notes.
Memory is private to each project. Audience levels decide what an outward-facing agent may read, and leaks are scanned for. Guardrails for secrets, card numbers and IBANs are built and off by default.
One operator, no separate organisations or role-based access. No certifications. A private repository. A company that runs itself is a goal.
What is ready and what is notThis is one operator's own working instance: the founder's, with test projects in it. These are not customers.
Snapshot of 4 November 2026, counted from the live database.
37,292 Scheduled runs, since 27 October 2026
2 more runs ended in another state.
Subscriptions are not counted in the spend. The failed runs are counted too.
Two things we did not find in the public docs of nine other products, and the rows where they lead.
Built, with a cascade that learns from past attempts.
Built. Approving each dynamic call is not supported yet.
Not found in the public docs we read, 9 products, checked 4 November 2026.
A hosted cloud, customers, certifications, a marketplace and a community, the number of connectors, telephony, and an open licence. We have none of these today, or only in part.
| Row | LOOPmKit | n8n | Zapier Agents | Dify |
|---|---|---|---|---|
| Agents trigger each other through records and events | ||||
| Automatic routing to the cheapest suitable model | ||||
| Hard budgets per run, agent and project that stop a run | ||||
| Trust levels: read, draft, act | ||||
| Isolated code execution | ||||
| Hosted cloud offer | ||||
| SOC 2 or other enterprise certification | ||||
| Customers |
Checked on 4 November 2026.
The full comparison, with sourcesAnd any MCP server or REST API you add by hand. Credentials sit in an encrypted vault.
Hosted or local. The router picks the cheapest model that fits the task.
Eight task agents and a set of employee roles to start from.
Test an agent the way you test code, with golden datasets and replay.
On your machines. LOOPmKit is a Docker Compose stack: the database, the memory service and your credentials stay where you run it, and credentials sit in an encrypted vault. Model calls go to the provider you configure, or stay local with a local model. Product telemetry is off by default.
More on safety and controlOutput is checked against a JSON Schema, with a retry and a fallback model. Risky actions wait as drafts for a person, an agent that fails three runs in a row pauses itself, and a failed run resumes from its checkpoint. The model can still be wrong, which is why the approval step exists.
More on safety and control25 provider presets: Anthropic, OpenAI, Gemini, DeepSeek, Mistral, Ollama, LM Studio and vLLM among them, so hosted and local models both work. The endpoints were probed; not every provider has been used in depth.
See every featureYou set what each connector allows (read, draft or act), budgets per run, per agent per day and per project per day, and policy rules where deny wins.
More on safety and controlPlenty. This is an early product with a single operator and a private repository. Some finished features are not switched on yet, and the goal of a company run by agents is not reached. The status page lists all of it.
What is ready and what is notWhat exists today, counted from the code and its notes. Nothing on this page is a forecast.
Teams are starting to give AI agents real work. The first problems they hit are cost that runs away, mistakes that reach a client and dependence on one model vendor. LOOPmKit is built around those three.
Vitalii Cherepanov builds and runs LOOPmKit. vbcherepanov.com
Stage: early. One operator, a private repository, no public release and no set terms. This page makes no claims about market size, users or revenue.
The product is at an early stage with a single operator. The repository is private and the terms are not set. If this fits a problem you have, book a short call.
A 20-minute call with the author. The repository is private, so a conversation is the way in.