Flowise alternatives after the repository archive

The maintained alternatives to Flowise are Dify, n8n and Langflow — and the reason to move is that the Flowise repository is archived and read-only, not that the software stopped running. Both halves of that sentence matter, because the wind-down did not shut everything down at once: the open-source project is finished, the published packages still install without a warning, and the hosted Cloud is still publishing prices with no end date attached. Which of those three facts applies to you decides whether this is an urgent migration or a planned one.

Two name collisions to clear first. flowise.io is a different company entirely — a work-management consultancy — so any pricing you find there describes that business. And Langflow is not Flowise: near-identical name, similar canvas, completely different ownership, licence and hosting story. Do not carry a fact from one to the other.

What actually shut down, and what did not

The vendor published a milestone table on its sunset page, and the stated reason is worth quoting rather than paraphrasing: "developers are increasingly relying on new coding agents such as Claude Code/OpenClaw to handle complex tasks. The typical rigid workflow low code approach quickly hits the limit when it comes to complexity." The milestones are July 29, 2026 for announcement and code freeze, repository archival in August, and August 31, 2026 for end of life of official core-team presence.

The archival date disagrees between two official sources. The sunset table says August 10, 2026; the "The Future of Flowise" announcement thread and the repository's own final push are both dated August 13, 2026. Use August 13 for your own records and note that the vendor page says otherwise.

One licence correction before anyone forks. The announcement says "the Apache 2.0 licensed code is yours to keep building on", but LICENSE.md carves out everything under packages/server/src/enterprise and files carrying an explicit copyright notice, which sit under a commercial licence requiring a valid Enterprise subscription for production use and forbidding redistribution. A fork is Apache 2.0 minus that directory, from a vendor that is winding down. No successor has surfaced under the project's own name: a GitHub repository search restricted to repositories with "flowise" in the name, pushed since August 20, 2026, turns up nothing above one star. A fork renamed to something else would not appear in that search, so read it as "no obvious successor", not as proof there is none.

Which alternative fits which Flowise workflow

Pick by the shape of what you already built, not by star count. All three below were checked on September 19, 2026 and all three are active, with pushes on September 18.

Langflow carries its own warning, and it is the same warning Flowise just delivered. It was acquired by DataStax in April 2024, IBM announced its intent to acquire DataStax in February 2025, and the hosted offering is already gone — a DataStax release note dated April 9, 2026 reads "Removed DataStax Langflow from Astra. As an alternative, you can use Langflow OSS." Today Langflow installs as a desktop app, Docker image, Python package or from source, and its site navigation has no pricing page. Choose it for the MIT-licensed code you will host yourself, not for a vendor to buy hosting from.

What migrates, what does not, and how to back out

Flowise's exporter source defines exactly fifteen record types: agent flows (both generations), four assistant kinds, chat flows, chat messages, message feedback, custom templates, document stores, document-store file chunks, executions, tools and variables. The routes expose export, import and chatflow-messages as permission-gated POST endpoints, so a whole-workspace dump can be scripted rather than clicked.

Three boundaries decide how long this takes.

• Credentials do not move. The official cloud-migration guide says so directly — "Credentials are not exported. You will have to create new credentials and use those in the flows and assistants" — and the exporter confirms it: the word credential appears zero times in that file. Every key, database password and OAuth token gets re-entered by hand.
• IDs survive, URLs change. The same guide notes that exported IDs stay the same, so an application calling a flow by ID only needs its base URL updated. That is the one thing that makes a Flowise-to-Flowise move cheap — and it buys you nothing moving to a different product.
• Nothing in those three repositories reads the format. A GitHub code search across langgenius/dify, n8n-io/n8n and langflow-ai/langflow returns zero hits for "flowise" in any of them, so none of the three ships an importer or converter. That search covers the indexed default branch, not third-party plugins. Use the export as a written specification and rebuild each flow.

Before any of that, snapshot the state. On a default self-hosted install the database is SQLite at ~/.flowise and blob storage at ~/.flowise/storage, per the environment-variable reference — but the credential encryption key lives outside the database, and without it those credentials are gone for good.

# Default self-hosted layout: the SQLite database and the blob storage
# folder both live under one directory. Snapshot it before you touch anything.
cp -a ~/.flowise ~/flowise-backup-$(date +%F)

# The encryption key is NOT inside that database. If it does not travel with
# the data, every stored credential is unrecoverable.
#   SECRETKEY_STORAGE_TYPE  local | aws
#   SECRETKEY_PATH          default: Flowise/packages/server
#   FLOWISE_SECRETKEY_OVERWRITE  replaces the file-based key entirely

# Pin the final release rather than tracking a floating tag, so a future
# registry change cannot move the image under you.
docker pull flowiseai/flowise:3.1.4

Rollback is then simple, which is the point of doing it first: keep the old stack cold rather than deleted until the replacement has run your real traffic for a full billing cycle, and rotate the keys the old stack held only after you decommission it.

What the replacements cost, and where the money actually goes

Two caveats on this table. Dify and n8n render only their annual-billing figures server-side; their monthly toggles are client-side and those numbers were not captured, so do not derive a monthly price from these. And the Flowise Cloud row is here to describe the market, not as a purchase recommendation — it is a priced tier on a site whose own banner announces a sunset.

Every one of those prices buys the builder. The model tokens are billed separately at whichever endpoint you point it at, and at real volume they dominate. Take the Starter tier's own published quota as the anchor: 10,000 answers in a month, each assumed to send 8,000 input tokens of prompt plus retrieved context and return 600 output tokens. Those assumptions are illustrative, not measured; rates are live Kunavo catalog prices per million tokens.

This is token arithmetic, not a measured task cost and not a bill ceiling. The reading that matters: under these assumptions the model line runs from roughly $44 to $132 a month against a $35 platform subscription, so "which builder is cheapest" is rarely the first question to answer. Model choice and prompt size move the bill by more than the platform does. One more line to budget separately: a retrieval app also embeds documents at ingest and queries at read, and Kunavo does not serve an embedding model — that step runs locally or against an outside provider and is billed there, not here.

Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. The minimum top-up is $10 in prepaid credit, which is a funding minimum rather than a task fee or a subscription — see billing details. Once you have chosen a builder, create a Kunavo account and follow the chat-completions reference for the base URL and key format. For the general pattern all three builders use, see OpenAI-compatible API; for the retrieval side, RAG implementation.

If you are staying on Flowise for now, fix the model list first

Flowise's OpenAI-compatible route is still there in the final release, and in source it carries no licence or plan check: the ChatOpenAI node ships in the Apache 2.0 packages/components tree, outside the packages/server/src/enterprise carve-out. On the ChatOpenAI node, open Additional Parameters and set Base Path (field name basepath, described as an override for the default base URL) and optionally Base Options, a JSON object of default headers. In the node source those map straight onto the client as baseURL and defaultHeaders. The API key is an ordinary ChatOpenAI credential.

The trap is the model name. On the plain ChatOpenAI node it is an async options field loaded by a listModels method that reads a models.json file — never your endpoint's own GET /v1/models. The default source is a raw file on the archived repository's main branch, so that dropdown is now frozen permanently: read on September 19, 2026 it tops out at claude-opus-4-7 for Anthropic and gpt-5.5 for OpenAI, and it will never gain a newer name. Setting Base Path alone leaves you choosing from a list that stopped moving in August.

Two documented ways out. Use the separate ChatOpenAI Custom node, whose Model Name is a free-text string field and which carries the same Base Path and Base Options; or set MODEL_LIST_CONFIG_JSON to your own file or URL, which replaces the dropdown source globally.

{
  "chat": [
    {
      "name": "chatOpenAI",
      "models": [
        {
          "label": "claude-sonnet-4-6",
          "name": "claude-sonnet-4-6",
          "input_cost": 0.0000012000,
          "output_cost": 0.0000060000
        },
        {
          "label": "claude-haiku-4-5",
          "name": "claude-haiku-4-5",
          "input_cost": 0.0000004000,
          "output_cost": 0.0000020000
        }
      ]
    }
  ]
}

Three boundaries on that route, all read from source on September 19, 2026. The Custom node calls a deny-list check against the base path, and the default deny list blocks, among others, localhost, 127.0.0.0/8, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 169.254.0.0/16, ::1 and fc00::/7 with "Access to this host is denied by policy" — a public HTTPS endpoint is unaffected, a proxy on localhost or a LAN address is not, and HTTP_SECURITY_CHECK and HTTP_DENY_LIST control it. The plain ChatOpenAI node does not run that check. And the Anthropic-Messages protocol has no base-URL field on the node that speaks it: the chat-model node directory contains one Anthropic-Messages node, ChatAnthropic, and its source declares no base-URL, base-path or API-URL input. So through that node an Anthropic-Messages endpoint is not addressable, and it has to be consumed through the OpenAI-compatible route instead — see the Anthropic base URL doc for where that protocol is available. One documentation quirk while you are reading: the official page titled "ChatOpenAI" is served from a URL slug containing "azure", and it is the canonical Base Path reference for the plain node, not Azure documentation.

A caveat on all of the above, stated the same way the AI agent API directory states it: Kunavo has not runtime-tested Flowise, Dify, n8n or Langflow against its endpoint. Every configuration claim here is transcribed from vendor documentation and repository source, not measured end to end. Keep a working route available, run one bounded task, and read the charge your own account recorded for it.

A staged cutover that you can reverse

1. Snapshot the data directory and the encryption key, as above, and pin the image to 3.1.4.
2. Export every workspace now, while the instance is up. If Cloud is your deployment, do this first — no export deadline is published, which is not the same as there being none.
3. Write down every credential and what it is for, before you rotate anything. None of them comes across.
4. Rebuild one representative flow by hand in the target, and only one. That is where you learn what the target cannot express.
5. Run both stacks on the same inputs and diff the outputs for a period you choose in advance.
6. Cut traffic over. Keep the old stack cold, not deleted, for one full billing cycle.
7. Decommission, then rotate the keys the old stack held.

If the conclusion of step 4 is that you do not need a visual builder at all — which is the reason the vendor itself gave for winding down — then the comparison worth reading is AI gateway and AI cost optimization rather than another canvas.

FAQ

Read directly for this page on September 19, 2026: the Flowise README, models.json, LICENSE.md, the ChatOpenAI, ChatOpenAI Custom and ChatAnthropic node sources, the export-import service source, the npm registry document, the Docker Hub repository document, flowiseai.com and flowiseai.com/sunset. Re-checked on September 21, 2026: the archive flag and final-push date, the announcement thread, the Dify, n8n and Langflow repository status, releases and licences, the Dify and n8n pricing pages, the DataStax Astra release note, the Flowise Terms of Service, and the Anthropic credential or component definitions in n8n, Dify and Langflow. Star counts were read on September 19 and move daily. Kunavo token rates are read from the live catalog, and every dollar figure in the worked example is illustrative token arithmetic.