Skip to content

Write the process down. Kinenix runs it.

Kinenix is an open-source RPA framework in Python. Each automation is a flow.md file that people and AI assistants can read, review in git, and run on any machine, with local language models for the steps that need judgment and no licence per robot.

pip install kinenix
Latest release
1.0.0
Licence
Apache-2.0
Runs on
Python 3.10+ on Windows, Linux, macOS
For AI agents
llms.txt
flows/bot_fx_rate/flow.md
# BOT FX Transfer Rate Collector
> Description: Collect Bank of Thailand transfer rates for one month, write a CSV, and email it.

### 3. Open BOT Exchange Rate Page (`web.open`)
- **url:** `${config.url}`
- **on_error:** retry
- **max_retries:** 2

### 18. Read Transfer Rates Per Currency (`logic.loop`)
- **items:** `${config.currencies}`
- **item_var:** `currency`

### 19. Write Grouped Rates To CSV (`csv.write`)
- **file_path:** `${config.output_path}`
kinenix run bot_fx_rate
  1. Completed [step_3] 'Open BOT Exchange Rate Page' in 2.41s
  2. Skipped [step_4] 'Accept Cookie Banner': optional, not shown
  3. Completed [step_18] 'Read Transfer Rates Per Currency' in 6.87s
  4. Completed [step_19] 'Write Grouped Rates To CSV' in 0.02s
  5. Completed [step_22] 'Email CSV Report' in 1.12s
  6. Saved JSON log to: logs/bot_fx_rate/…/success.json

Built for automation you have to keep running

Robots are easy to start and hard to maintain. Kinenix is designed around the second part.

No licence per robot

Kinenix is Apache-2.0. Run unattended workers on the VMs, office PCs, or Raspberry Pis you already have. Excel and CSV work through openpyxl and pandas, so no Office installation is needed either.

Processes as documents

A flow is Markdown that an analyst can read and a reviewer can approve in a pull request. Describe the task in requirements.md and an AI assistant writes the flow.md for you.

Local AI, private data

ai.extract, ai.prompt and ai.decide call small models through Ollama on your own hardware. Invoices, HR records and customer e-mails never leave your network, and there are no API tokens to pay for.

Failures you can explain

Retries, On Error and Finally sections, one record per run with a screenshot of each failure, and secrets redacted before anything is logged. The Hub summarizes what went wrong in plain words.

From a written task to a recorded run

  1. Describe the task

    Create a project and write what the robot should do in plain language.

    kinenix init "Get stock data"
  2. Check the flow

    Validation reports unknown actions and expression errors with their line number.

    kinenix validate
  3. Run it, or schedule it

    Run once from the terminal, or hand it to a worker with a schedule or a file watch.

    kinenix run
  4. Review every run

    Workers report to a self-hosted Hub you can query with SQL from the command line.

    kinenix hub query

AI where judgment is needed, rules everywhere else

Steps run in the order written, the same way every time. An AI step is one action among the others, with a defined output, so a model can classify an e-mail or read an invoice without deciding what the robot does next.

The format itself was chosen by testing. Small local models wrote the same flow in flow.md and in YAML; flow.md came out ahead for every model of 3B parameters or more.

5 of 5 flows parsed in flow.md, against 2 of 5 in YAML, written by qwen2.5-coder:7b

Read the benchmark
### 2. Route Customer E-mail (`ai.decide`)
- **state:** ${email.body}
- **question:** {"name": "intent",
    "type": "choice",
    "options": ["Receipt request",
                "Problem report",
                "Price enquiry",
                "Cancellation"]}
- **output_var:** `route`

### 3. Open A Support Ticket (`http.request`)
- **condition:** ${route.choice} == Problem report

How it compares

Kinenix sits between a commercial RPA suite and a folder of Python scripts.

KinenixCommercial RPA suitesPlain Python scripts
Cost of one more robotNonePer-robot or per-user licenceNone
How a process is writtenMarkdown, diffable in gitProprietary designer filesPython code
AI stepsLocal models via Ollama, built inUsually a cloud add-onWire up an SDK yourself
Where documents goStay on your machinesOften through vendor cloud servicesStay on your machines
Unattended runsWorker with schedules and file watchOrchestrator, licensed separatelycron or Task Scheduler
Run historySelf-hosted Hub with SQL viewsVendor orchestratorBuild it yourself
MaturityBeta; APIs still changeMature, with vendor supportUp to you

What is ready today

Kinenix 1.0 is stable: a flow written for 1.0 keeps running on every 1.x version. Pin the version a flow needs with > Requires: kinenix>=1.0.0 so older workers refuse it instead of misreading it.

  • Ready

    kinenix-core

    Flow interpreter, 47 actions across web, Excel, CSV, files, HTTP, e-mail and AI, and the kinenix CLI.

  • Ready

    kinenix-worker

    Unattended runner with schedules and file-watch triggers. Dispatch from the Hub is planned.

  • Early

    kinenix-hub

    Central run records, dashboard, and plain-language failure summaries.

  • Paused

    kinenix-studio

    Visual flow editor. Flows are written with an AI assistant instead for now.

See the roadmap

Run your first flow in about ten minutes

The quickstart installs Kinenix, runs an example, and connects a worker to the Hub.

Start the quickstart