LOOPmKit and Dify, side by side
Dify combines a canvas for workflows and agents with a marketplace and a cloud. Here is how its public pages compare with our own feature list.
Checked on 4 November 2026. We read the product pages on that date. Sources are linked next to the notes.
What each one is
LOOPmKit
A self-hosted builder of autonomous AI agents. Agents wake one another through records, every run has budget limits, and risky actions wait for a person to approve them. It is an early product with one operator and a private repository.
FeaturesDify
Dify, from LangGenius, has a visual canvas for workflows and agents, a Community Edition that runs with Docker Compose, a marketplace for plugins and templates, and a hosted cloud. Its license is described as Apache 2.0 with additional conditions, and the Community tier is aimed at non-commercial use. A Human Input node lets a reviewer edit content and choose a branch.
Dify product pageWhere Dify leads
Rows where Dify is marked as found and LOOPmKit is not fully there.
- Self-hosting and licenseCommunity Edition via Docker Compose. Dify Open Source License, described as Apache 2.0 with additional conditions; Community tier aimed at non-commercial use. Source
- Hosted cloud offerDify Cloud, Sandbox free to Team 1,590 USD per year. Source
- Ready marketplace and community sizeAbout 160.4k GitHub stars; Dify Marketplace for plugins and templates. Source
What we did not find in the Dify docs
Rows where LOOPmKit has it built and the public pages we read for Dify do not say. Not found in the public docs we read is not the same as no.
- Agents trigger each other through records and eventsTriggers on a record, on another agent finishing, on a schedule or a webhook; also an agent as a tool of another. The live call of an agent as a tool awaited a key.
- Automatic routing to the cheapest suitable modelAuto mode picks the cheapest model that meets the tools, vision, context and quality needed, and a learned cascade starts on proven models.
- Hard budgets per run, agent and project that stop a runPer run, per agent per day and per project per day; a run over the cap is skipped, and parallel calls reserve budget first.
- Built-in savings: deduplication, response and prompt cacheThree savers on by default: dedupe by hash, an exact response cache and the provider prompt cache. No semantic cache.
- Trust levels: read, draft, actThe effective level is the lower of the action and the connector scope. Dynamic Gmail calls need act, since per-call approval is not supported yet.
- Built-in persistent memoryA separate memory service from the same author in the same stack; the run goes on if it is down. The "memory required" policy is not rolled out.
- Evals and golden datasets10 kinds of evaluators, golden datasets, replay and comparison of versions and models. The prompt optimiser is not rolled out.
- Isolated code executionPer-job containers with a read-only root, dropped capabilities, limits and an egress allowlist. Component tests only; no run with a real vendor subscription yet.
- Company wizard with roles10 editable roles, up to 32, with budgets and colleague-call rules in one transaction. Checked in a browser with a scripted model. The goal of a company that runs itself is not reached.
Found in both
Rows marked as found for LOOPmKit and for Dify.
Visual canvas.
When to choose which
Choose Dify when
You want a canvas for workflows and agents together with a large community (about 160.4k GitHub stars), a marketplace for plugins and templates, and a hosted cloud.
Choose LOOPmKit when
You need budgets per run, agent and project, agents that wake one another on records, and trust levels. Dify plans have message-credit quotas, which is not the same as a budget that stops a run.
Read this with care
The LOOPmKit column marks what is built but not rolled out, or tested only against stand-ins. A question mark for the other product means the pages read do not say, not that it is missing. Only one operator has run LOOPmKit so far.
All rows and notes on the LOOPmKit side What is ready and what is not
Row by row
Each row shows LOOPmKit and Dify. Our notes come from our own feature inventory.
- Found
- Partly
- No
- Unknown: the pages read do not say
-
Self-hosting and license
LOOPmKit. A Docker Compose stack for one operator. The repository is private and has no license, so it is not open source and others cannot self-host it today.
Dify. Community Edition via Docker Compose. Dify Open Source License, described as Apache 2.0 with additional conditions; Community tier aimed at non-commercial use. Source
-
Visual canvas
LOOPmKit. A canvas of agents and gates.
Dify. Visual canvas for workflows and agents. Source
-
Agents trigger each other through records and events
LOOPmKit
Dify
LOOPmKit. Triggers on a record, on another agent finishing, on a schedule or a webhook; also an agent as a tool of another. The live call of an agent as a tool awaited a key.
Dify. No record-event agent chaining doc was checked; other forms of chaining may exist.
-
Model providers, local models included
LOOPmKit. 25 provider presets including Ollama, LM Studio and vLLM. The addresses were probed with invalid keys only.
Dify. Hundreds of models and any OpenAI-compatible API; local models not explicitly stated on the page read. Source
-
Automatic routing to the cheapest suitable model
LOOPmKit
Dify
LOOPmKit. Auto mode picks the cheapest model that meets the tools, vision, context and quality needed, and a learned cascade starts on proven models.
Dify. Not found in the vendor docs read.
-
Hard budgets per run, agent and project that stop a run
LOOPmKit
Dify
LOOPmKit. Per run, per agent per day and per project per day; a run over the cap is skipped, and parallel calls reserve budget first.
Dify. Plans have message-credit quotas; not a per-run or per-agent budget.
-
Built-in savings: deduplication, response and prompt cache
LOOPmKit
Dify
LOOPmKit. Three savers on by default: dedupe by hash, an exact response cache and the provider prompt cache. No semantic cache.
Dify. Not stated on the vendor pages read; not assumed to be absent.
-
Approval queue with editing of arguments
LOOPmKit. Drafts wait in a queue, the approver can edit the arguments, and disclosure is checked again after the edit.
Dify. Human Input node: reviewer edits pre-filled content and chooses a branch; this is content review, not tool-argument editing. Source
-
Trust levels: read, draft, act
LOOPmKit
Dify
LOOPmKit. The effective level is the lower of the action and the connector scope. Dynamic Gmail calls need act, since per-call approval is not supported yet.
Dify. No read/draft/act ladder found in the docs read.
-
Built-in persistent memory
LOOPmKit
Dify
LOOPmKit. A separate memory service from the same author in the same stack; the run goes on if it is down. The "memory required" policy is not rolled out.
Dify. Knowledge bases documented; agent memory not confirmed.
-
Evals and golden datasets
LOOPmKit
Dify
LOOPmKit. 10 kinds of evaluators, golden datasets, replay and comparison of versions and models. The prompt optimiser is not rolled out.
Dify. Retrieval testing only on the page read.
-
MCP as client and as server
LOOPmKit. As a client, any stdio or HTTP MCP server. As a server, 11 tools; built and not rolled out.
Dify. MCP servers as tools (HTTP transport only); publishing as MCP server referenced but not read. Source
-
Browser and desktop control
LOOPmKit
Dify
LOOPmKit. Headless Chromium in the API image; a Mac bridge with manual setup; logged-in browser profiles are in progress.
Dify. Not stated on the vendor pages read; not assumed to be absent.
-
Isolated code execution
LOOPmKit
Dify
LOOPmKit. Per-job containers with a read-only root, dropped capabilities, limits and an egress allowlist. Component tests only; no run with a real vendor subscription yet.
Dify. Not checked on the pages read.
-
Company wizard with roles
LOOPmKit
Dify
LOOPmKit. 10 editable roles, up to 32, with budgets and colleague-call rules in one transaction. Checked in a browser with a scripted model. The goal of a company that runs itself is not reached.
Dify. Not stated on the vendor pages read; not assumed to be absent.
-
Number of connectors
LOOPmKit
Dify
LOOPmKit. 61 presets and any MCP server or REST API. None is checked live one by one.
Dify. Marketplace exists; no count read.
-
Telephony
LOOPmKit
Dify
LOOPmKit. An inbound Twilio receptionist, tested against a simulated Twilio and off by default. No real call and no outbound calls.
Dify. Not stated on the vendor pages read; not assumed to be absent.
-
Hosted cloud offer
LOOPmKit. Self-hosting only is documented.
Dify. Dify Cloud, Sandbox free to Team 1,590 USD per year. Source
-
Ready marketplace and community size
LOOPmKit. A private repository and no public community. 23 templates inside the product.
Dify. About 160.4k GitHub stars; Dify Marketplace for plugins and templates. Source
-
SOC 2 or other enterprise certification
LOOPmKit. No certification.
Dify. SOC 2 and ISO 27001 badges on pricing page without audit details. Source
-
Customers
LOOPmKit
Dify
LOOPmKit. No external customers and no revenue yet.
Dify. No customers named on pricing page.
Checked on 4 November 2026. Notes on other products are quoted in English from the pages read.
Not affiliated
LOOPmKit is not affiliated with, endorsed by or sponsored by LangGenius. Dify and other product names belong to their owners and are used here only to name the products compared.
Other comparisons
LOOPmKit vs n8nLOOPmKit vs CrewAILOOPmKit vs LangGraphLOOPmKit vs Zapier Agents The full comparison of all nine products
Keep reading
- Every feature, grouped and marked
The full list of what LOOPmKit does today, with the real status of each item.
- Budget limits, trust levels and human approval
How agents are kept inside limits you set.
- What is ready and what is not
A plain status page with the real limits of an early product.
- Roles in an AI agent team
What every role does, who it hands work to and what needs your approval.
Early access starts with a call
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.