Agent orchestration
Coordinate planners, workers and tools as one flow, with clear hand-offs and retries.
Learn more
Open-source · AI infrastructure
Orchestrate AI agents. Ship LLM pipelines that flow.
An open-source, high-performance framework for building, chaining and running autonomous AI agent workflows.
Build
0xfaster pipeline setup (goal)
0%open source
AnyLLM provider (by design)
0core modules
Targets we're building towards, not benchmarks yet.
Why NeuralFlow
One framework from first agent to production pipeline.
LLM pipelines
Chain load, prompt, validate and save steps into pipelines that batch and recover.
Learn moreOpen source first
Built in the open: read the code, run it anywhere, shape where it goes next.
Learn moreTry it
Watch a flow run, node by node.
A mini demo of how a request moves through a NeuralFlow graph.
01InputUser request
02Planner AgentBreaks down the task
03Tool Callsearch_docs()
04OutputFinal answer
- Press “Run flow” to start.
Product
Four modules. One flowing API.
Orchestrate agents, build pipelines, see every step and extend it all with plugins.
Agents that work together
Define agents with a model and tools, then chain or branch them into a flow that plans, acts and answers.
- Planner, worker and reviewer patterns
- Typed tools from plain Python functions
- Retries, timeouts and hand-offs per step
from neuralflow import Flow, Agent, tool
@tool
def search_docs(query: str) -> list[str]:
return index.search(query, k=5)
planner = Agent("planner", model="any-llm", tools=[search_docs])
writer = Agent("writer", model="any-llm")
flow = Flow("answer-question")
flow.chain(planner, writer)
result = flow.run("How do refunds work?")
print(result.output)Pipelines for real data
Compose load, prompt, validate and save steps into a pipeline that runs in batches and picks up where it failed.
- Batching and concurrency controls
- Schema validation on every LLM output
- Automatic retries and resumable runs
from neuralflow import Pipeline, steps
pipe = (
Pipeline("enrich-leads")
.add(steps.Load("leads.csv"))
.add(steps.LLM("Summarise each company", model="any-llm"))
.add(steps.Validate(schema=LeadSummary))
.add(steps.Save("warehouse.leads"))
)
pipe.run(batch_size=32, retries=2)See every step
Every run produces a trace: which agent did what, which tools it called, how long it took and what it cost.
- Step-level traces with latency and tokens
- Export to console, files or your own stack
- Replay a run to debug it
from neuralflow import Flow
from neuralflow.observe import Tracer
flow = Flow.load("answer-question")
tracer = Tracer(export="console")
with tracer.session("debug-run"):
run = flow.run("Draft a reply to ticket #4821")
for span in run.trace:
print(span.step, span.latency_ms, span.tokens)Extend anything
Plugins hook into the flow lifecycle, so you can add tools, providers, storage or alerts without forking.
- Lifecycle hooks for steps and runs
- Bring your own model provider
- Share plugins as regular packages
from neuralflow import Flow
from neuralflow.plugins import Plugin, register
@register
class SlackNotify(Plugin):
name = "slack-notify"
def on_step_end(self, step, result):
self.post(f"{step.name} finished in {result.ms} ms")
flow = Flow.load("enrich-leads", plugins=["slack-notify"])How it works
From idea to running flow in three steps.
- 01
Define your agents and tools
Pick a model, describe the role, hand it plain Python functions as tools.
- 02
Connect them into a flow
Chain, branch or loop agents and pipeline steps into one graph.
- 03
Run, monitor and iterate
Run locally or at scale, read the trace, then tune prompts and steps.
Under the hood
Built for the way developers actually ship.
- ⚡
Built for performance
Async by default, with batching and concurrency where it counts.
- ◎
Model-agnostic
Designed to work with any LLM provider, hosted or local.
- { }
Developer-friendly API
Small, typed, readable. Flows are just Python.
- ◉
Observability built in
Traces, timings and token counts for every step, out of the box.
- ✚
Extensible with plugins
Hook into the lifecycle to add tools, providers and integrations.
Want to build with it first?
Get early accessUse cases
What teams will build with NeuralFlow.
Filter by area, then open a card to see an example workflow.
Customer support automation
Triage tickets, draft grounded replies and escalate what needs a human.
- Classify the incoming ticket by intent and urgency.
- Retrieve matching help-centre articles and past replies.
- Draft a reply grounded in those sources.
- Route: send low-risk replies for quick approval, escalate the rest.
Research & document analysis
Read long PDFs and reports, extract facts and write cited summaries.
- Ingest PDFs and split them into sections.
- Extract key claims, figures and dates per section.
- Cross-check claims across documents with a reviewer agent.
- Summarise into a brief with citations back to pages.
RPA-style business process automation
Let agents handle the judgement steps between your existing systems.
- Watch an inbox or folder for new invoices.
- Read each invoice and extract vendor, amount and due date.
- Validate against purchase orders and flag mismatches.
- Post approved entries to the finance system and log the run.
Internal developer copilots
Answer questions about your codebase, runbooks and internal APIs.
- Index repos, docs and runbooks on a schedule.
- Plan: the planner agent decides which sources a question needs.
- Search code and docs with tool calls.
- Answer with snippets and links to the exact files.
Data pipeline enrichment
Add LLM steps to data pipelines: classify, tag, summarise and validate rows.
- Load a batch of rows from the warehouse.
- Enrich each row with an LLM step (category, summary, tags).
- Validate outputs against a schema; retry failures.
- Write enriched rows back and record a trace per batch.
Sales & CRM assistants
Research leads, prep account briefs and keep the CRM up to date.
- Pick up new leads from the CRM.
- Research each company from public sources.
- Write a short account brief and a suggested first message.
- Update the CRM record and notify the account owner.
No use cases in this filter yet.
About
Our mission
Make building with autonomous AI agents as simple and reliable as building a web app — in the open.
Values
What we hold ourselves to.
Open by default
The code, the roadmap and the decisions happen in public.
Developer first
Small APIs, clear errors, good docs. If it's awkward to use, it's a bug.
Built to scale
The flow you prototype on a laptop should be the flow you run in production.
Founder
Who's building it.
Amith Viswas Reddy Dwarampudi
Founder
Speaks English, Hindi, Japanese and Telugu.
Connect on LinkedInRoadmap
Where NeuralFlow is headed.
Honest stages, no dates we can't keep.
Now In progress
Core framework in active development
Flows, agents, tools and pipelines: the core API is being built and tested.
Next
Public open-source release + documentation
The repository goes public with guides, examples and an API reference.
Later
Community plugin marketplace
A place to find and share plugins for tools, providers and integrations.
Future
Hosted tooling & dashboard
Optional hosted tools to run, trace and manage flows from one dashboard.
FAQ
Questions, answered.
Yes. NeuralFlow is open source and free to use.
It's designed to be model-agnostic, so you can plug in the provider or local model you prefer.
Follow the roadmap above, or join early access to hear first.
Yes. Community contributions are welcome once the repository is public.
Use the contact page to request it.
Contact
Get early access.
Tell us what you'd build. We'll be in touch as the release gets closer.
Follow the founder for build updates.
Open LinkedIn