Turn a business decision into a tested decision engine.
AutoJEV is Jev in self-service mode. Tell your coding agent the decision you need to make. It works out what kind of decision it is, shows the alternatives it weighed, and builds the engine in your project.
Works with Claude Code, any MCP client, or any agent with a shell. Plain Python, no dependencies.
6/7 agree (86%; need 75%)
VERIFIED
The hard part
Calling Jev is easy. Framing the decision is not.
Jev answers three kinds of question. For a vague business problem, someone has to decide which kind it is, how many separate questions it needs, where the thresholds go, and what happens when Jev is not sure.
Is this true?
Is this message spam? Is this claim over policy?
Returns yes / no
Which one?
Which desk should this message go to?
Returns one of N options
Where on a spectrum?
How urgent is this ticket, from 0 to 1?
Returns a number with a threshold
“Every inbound member message should go to the right desk, and anything risky should be flagged.”
One multiple choice? Five yes/no flags? A score? A pipeline of all three? Most people guess, then find they cannot tune it. AutoJEV makes that call explicit, shows its reasoning, and only then writes code.
How it works
From a sentence to a tested engine
- 1
Describe the decision
Say what you need decided, in plain words, to the coding agent you already use.
- 2
Review the designs it weighed
It reads your project, asks at most three questions, and ranks candidate designs, the naive one included, before writing anything.
- 3
Get a tested engine
The engine lands in your project with its spec, tests, sample cases and a record of every trade-off.
- 4
Keep it current
As your project changes, it proposes updates through the spec and rolls back on failure.
What you get
Plain Python you can read, run and change
Every engine is a folder in your project. The spec is the source of truth, the tests need no key, and the decision record explains why this design beat the others.
- engine.py
- Questions, thresholds, policy and a CLI
- spec.json
- The source of truth. Edit it and rebuild
- decision_record.md
- Designs considered, scores, assumptions, risks
- samples.json
- Cases, with the outcome each should reach
- test_engine.py
- Behavioural tests, no key needed
- verification.json
- Receipt from the live check, tied to this exact spec
Trust and safety
Generated code deserves suspicion, so these are deliberate
Your key stays out of the chat
Entered at a hidden prompt in your own shell and saved to a git-ignored file. No tool accepts it.
Generated code is not trusted
Text from a spec is escaped before it reaches code, so a poisoned project file cannot become executable.
It says what it has not verified
Engines read NOT VERIFIED against Jev until the live check has run, and a custom endpoint is labelled as one.
Changes are bounded
Recalibration edits only through the spec, rolls back on failure, and proposes rather than applies a change of design.
What was tested, honestly. The live code paths were tested against a local stand-in for the Jev API, not the production API. That proves the plumbing. Nothing in it says how Jev will judge your data, so your first live check is the first real test.
Quick start
Two commands, then describe a decision
Install the Claude Code plugin, then ask in any project. Building and testing work offline. For the live check you add a Jev API key at a hidden prompt in your own shell.
Not on Claude Code? It also runs as an MCP server or a plain command line. See the README.
New to Jev? Try the Jev Playground first.
# 1. install
claude plugin marketplace add exponen-agi/autojev
claude plugin install autojev@autojev
# 2. in any project, describe the decision
/autojev should we auto-approve this expense claim,
or send it to a manager?
# 3. run what it built
python decision_engines/expense-review-gate/engine.py --demoQuestions
Before you try it
What is AutoJEV?
AutoJEV is Jev in self-service mode for business domains. You tell your coding agent the decision you need to make, and it builds a tested Jev decision engine in your project. Jev answers three kinds of question: is this true, which one, and where on a spectrum. AutoJEV works out which one your problem is.
Do I need a Jev API key?
Not to build or test. Engines are built and tested offline without a key, and they say they are not verified against Jev. You need a key to run the live verification step. The key is entered at a hidden prompt in your own shell and never goes through the chat.
Which agents does it work with?
It installs as a Claude Code plugin. It also runs as an MCP server for any MCP client, and as a plain Python command line for any agent that can use a shell. It is plain Python 3.10+ with no dependencies.
Is it production ready?
This is the initial release. The live code paths were tested against a local stand-in for the Jev API, not the production API, so your first live verification is the first real test of how Jev judges your data.
Is it free?
AutoJEV is open source under the MIT licence. Jev API usage is billed by its provider, separately.
Make your business AI-native, starting this month.
You speak to the senior lead who would do the work. If we are not the right fit, we will say so.