checking the server…

Decisions,
not prose.

Ask Dev Decide a question about your code or your issues. It answers with a typed choice, a confidence and where the answer came from: from your own trained models in milliseconds, from Claude only when it must.

Try it live Install the plugin

An open-source, Jev-style decision engine: typed answers with a confidence, not chatbot prose, running on your own machine.

1Pick a decision2Paste your input3Branch on the answer
Example output
›
REVIEW
The highlighted tier is the one that answered. Every Claude answer trains the tiers above it.
0.68 ms
per local decision
80%
answered locally in steady use
90.6%
held-out routing accuracy
$0
for every local answer
Measured figures from the project's README (2-core VM; steady-usage simulation; held-out test set).
Live console

Ask it something.

Paste an issue, a request, an error or a diff. You get typed answers, each with how sure it is and where it came from. Local answers are instant and free; Claude is asked only when you allow it.

decide
connecting
Access token
Sent in a header with each request and used for that request only. Never stored by this page or the server.
1What should it decide?
2Your input
Try an example:
3How should it answer?
Ctrl + Enter
Your decision appears here
with how sure it is and where each answer came from.
local Claude rough guess
Checking rules
Asking the learned model
Recalling similar inputs
Anything still unsure goes to Claude (if allowed)
Your recent decisions
Kept in this browser only, never sent anywhere.
How it works

The fastest answer it can trust.

Every question walks down the tiers and stops at the first one that has earned trust for it. Most stop in the first three, on your own machine, for free.

01

Rules

Exact facts: a live API key, a migration file, docs-only changes.

~3 msfree
02

Learned model

Your trained model, used only above the confidence it proved on held-out data.

<1 msfree
03

Memory

Your own corrections, and inputs like ones already answered.

~1 msfree
04

Claude

One lean call for whatever is still open, with the local findings as evidence.

2–4 s~$0.001
05

Stronger model

Unsure answers to the decisions that matter, capped per day.

+5 s~$0.02

It gets faster and cheaper the more you use it.

Each answer Claude gives becomes a label. The local model trains on it, and it only replaces the current one after proving it is better on your own decisions. It can always roll back.

80% of questions answered locally in steady use (README simulation).
Decisions

One call, a whole judgment.

Bundles ask several typed questions at once and apply a pass, review or block gate. Pick one to try it with an example.

Pull requests

Risk, breaking changes, secrets, security, tests and migrations, gated.

pr :: DIFF

Issue triage

Type, domain, severity, effort and the model tier it needs.

triage :: <issue>

Routing

Which of your agents or skills should take a request.

route :: <task>

Incidents

Error category, severity, production impact and flakiness.

incident :: <log>

Deploys

Risk, rollback safety, migrations and performance sensitivity.

deploy :: DIFF

AI changes

Prompt injection, agency, evals, unpinned models, PII and cost.

ai_change :: DIFFInstall
Install

Two commands in Claude Code.

It runs on your machine and asks Claude with your own sign-in or your own ANTHROPIC_API_KEY. Nothing is shared.

Starter models, a documented settings file and the routing hint are included.

/plugin marketplace add pawanptu/dev-decide
/plugin install dev-decide@dev-decide
/dev-decide pr :: DIFF

For the HTTP API on your computer, or to choose whether to add the routing hook.

git clone https://github.com/pawanptu/dev-decide
python3 dev-decide/dev-decide-v5/install.py --router-hook

See which agent or skill was suggested for which task, what Claude really used, and how long each step took.

/plugin install dev-decide-graph@dev-decide
/dev-decide-graph
Or open http://127.0.0.1:47321/graph while Claude Code runs.

Private by design

  • Decides locally first; only unsure questions reach Claude.
  • Logs predictions only, never your input text.
  • Learned data stays in your repo's git-ignored .claude/dev-decide/.
  • Switches models only after a challenger proves better on your own decisions, and can roll back.
  • Standard-library Python. No pip installs.

Stop parsing prose. Branch on answers.

Open source, MIT licensed, and free to run locally.

Try it now View on GitHub