Flowise
Flowise is a visual builder for LLM applications. Flows are stored as a node graph and executed server-side, so the product's real behaviour lives in a graph rather than in a repo, and reading that graph is a genuine skill rather than an inconvenience.
Lead with this, it changes most answers
The project was archived on 13 August 2026 and is read-only. The final release is 3.1.4
(29 July 2026). The core team leaves Discord and GitHub on 31 August 2026, and the security
policy was replaced with a notice that vulnerability reports are no longer accepted.
Three consequences worth stating before giving anyone advice:
- Pin
3.1.3, not3.1.4. The final release ships a Docker image reported not to boot on a fresh volume (EACCESon/root/.flowise, missing@smithy/eventstream-codec,this.db.exec is not a function). It will never be fixed.3.1.3(25 June 2026) is the last usable release. - "Just fork it, it's Apache 2.0" is not accurate. Since
3.0.1the repo is dual-licensed:packages/server/src/enterprise/**andIdentityManager.tssit under a FlowiseAI commercial licence.IdentityManager.tsis load-bearing for auth and route mounting, so a fork intended to carry auth needs that resolved first. The README still says plain Apache 2.0;LICENSE.mdis the operative document. - Nothing marked "will be removed in a future release" ever will be. 46 nodes carry a DEPRECATING badge, including the whole LlamaIndex integration and the legacy agents. They are now permanent. That cuts both ways: no forced migration, and no fixes either.
Being frozen has one upside worth using: the facts below stop moving, so they can be stated precisely rather than hedged.
Work out which question you are answering
| The user wants to | Read |
|---|---|
| Understand what an existing flow actually does, or extract its prompts and logic | references/audit.md |
| Work out why flow A works and flow B does not, and what to copy | references/transfer.md |
| Design or modify a flow: node choice, branching, state, structured output | references/agentflow-v2.md |
| Call a flow from code, stream it, embed it, or move it between environments | references/api.md |
| Deploy, upgrade, scale, secure, or diagnose an instance | references/operate.md |
Read the one that fits. They are written to stand alone, so pulling in all five wastes context.
Reach for the client before writing throwaway code
scripts/flowise.mjs is a dependency-free Node client for reading an instance. Use it rather
than curling the API by hand, because it already absorbs the three encodings that sit between
you and the meaning of a flow.
export FLOWISE_API_ENDPOINT=https://flowise.example.com FLOWISE_API_KEY=...
node scripts/flowise.mjs flows # every flow: type, node count, id
node scripts/flowise.mjs nodes "My Flow" # node map with kinds, sizes, edges
node scripts/flowise.mjs vars "My Flow" # every {{ }} reference, per node
node scripts/flowise.mjs prompts "My Flow" # all prompts, HTML decoded
node scripts/flowise.mjs code "My Flow" --node "Search" # JS body + its inputs
node scripts/flowise.mjs flow "My Flow" --out flow.json # whole flow
node scripts/flowise.mjs get /api/v1/nodes # any endpoint, raw
node scripts/flowise.mjs diff a.json b.json # semantic diff, ignoring ids and canvas positions
Credentials come from --endpoint/--key, then the environment, then ~/.flowise.json.
Flow lookup accepts an id, an exact name, or a case-insensitive substring.
diff is built for the question "has production drifted from the flow I read on dev?" The same
flow on two instances has different node ids, credential ids and canvas coordinates, so a naive
comparison marks every node changed and answers nothing. This one normalises those away, compares
each node's inputs, and names the fields that actually moved. "No semantic differences" is a real
answer you can act on.
vars is the underrated one. A flow's state keys are named by whoever understood the problem,
so listing them and which nodes touch them explains the design faster than reading any prompt.
Five facts that decide most answers
Flows are one table, discriminated by type. CHATFLOW, AGENTFLOW, MULTIAGENT,
ASSISTANT. The trap: MULTIAGENT means Agentflow V1, not "a multi-agent flow". Agentflow
V2 is AGENTFLOW. Filter with /api/v1/chatflows?type=AGENTFLOW.
flowData is a JSON string inside the flow object. Parse it or the flow looks empty. This
is the single most common reason a first attempt returns nothing.
There is no /api/v1/agentflows endpoint. That path returns the single-page app's HTML with
a 200, which looks like a working response and is not. Agentflows come from the chatflows route
with a type filter. The bundled client detects this and says so.
$flow.state keys must be declared in the Start node. Operational nodes can update existing
keys but cannot create new ones, and state lives for exactly one execution. A flow that seems to
lose data between nodes is usually missing a declaration, not misusing the syntax.
overrideConfig is disabled by default since 2.1.4, per property, in the flow's Security
tab. An override that silently does nothing is nearly always this rather than a malformed body.
How to be useful here
Read the graph before believing the description. The person who built a flow will describe what they intended, accurately as they remember it, and be wrong in specific ways that cost real time. Flows are edited in a GUI with no version history a reader can consult, so the graph is the only account that is current. This matters most when someone asks you to copy what another flow does: verify the mechanism, then ask the human for the reasoning, which is the part the graph genuinely does not hold.
Prefer citing a node to asserting a behaviour. "Node 4 reduces each candidate to id, title and type before curation" is checkable. "The flow only passes titles" is an opinion until it names where. Analyses from this work usually end up in front of people making scheduling decisions, so they need to survive someone opening the flow.
Distinguish what is broken from what is unfixable. Since the archive, "file an issue" and "wait for the fix" are no longer available moves. When you hit a known bug, say whether a workaround exists, whether pinning an earlier version avoids it, and that upstream will not resolve it. That changes what the user should plan, which is the actual question behind most bug reports.
Say when the docs are wrong. The published documentation contains real defects: a plural
/predictions/ path that does not exist, a Docker example carrying another product's port and
an invalid DATABASE_TYPE, a Chatflow.type enum missing AGENTFLOW entirely, and a
WORKER_CONCURRENCY default of 10000 that is a copy error. references/operate.md and
references/api.md list these. Repeating a doc defect confidently is worse than saying you are
unsure.