mcts-agent
Coding agent that runs discriminative Monte Carlo Tree Search, with Gemini proposing actions and Jev Choice and Score guiding the search.
Overview
mcts-agent is an autonomous reasoning and execution agent built on discriminative Monte Carlo Tree Search. Gemini Flash proposes candidate actions for expansion, while Jev provides the fast structured evaluations: a Choice assigns prior probabilities over proposed actions, a Score evaluates hypothetical states as the value function, and a Choice also picks expansion width dynamically from small prime numbers based on uncertainty and goal complexity. Low early scores do not prune proposed paths.
The agent follows a closed loop of plan and choose, execute, review, adapt and assess. Only the immediate winning action is executed in the workspace, and git diffs and terminal observations are folded back into context for the next decision. After execution, surviving grounded paths are rescored, and the tree is rebuilt only when no actionable path reaches a reuse threshold. A local web UI lets you prompt the agent, follow its output and changed image files live, and load saved run logs. It is MIT licensed with CI.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What prior probability should each proposed action get? | — |
| Score | How promising is this hypothetical state? | — |
| Choice | How wide should this node expand? | 2 · 3 · 5 · 7 · 11 · 13 |
How it works
- Gemini proposes candidate actions
- Jev assigns priors and scores states during PUCT search
- Execute only the best immediate action
- Fold diffs and terminal output back into context
- Rescore surviving paths or rebuild the tree
Strengths
- Combines generative proposals with fast evaluation
- Only one action executed per step
- Tree reuse reduces recomputation
- Live web UI and run logs
Caveats
- Requires both Gemini and Jev
- Many Jev calls per search
- Research-style agent
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- agent with local web UI
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Discriminative Monte Carlo Tree Search using TypeSafe Jev System One Primitives and Gemini.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev do in the search?
It provides action priors with Choice and state values with Score.
Who proposes actions?
Gemini Flash.
How many actions run at once?
Only the single best immediate action per step.
Related projects
One skill and MCP server for Claude Code, Codex and Hermes where Jev gates large-model turns, routes models and guards Git steps.
MCP serverbetaPythonPython harness for open-ended goals: AI-authored tools and checks, Jev choosing the next tool and arguments from fresh state.
ProjectbetaPythonPython agent runtime where Jev Choice decisions pick tools and arguments, and low-confidence picks go to an LLM for review.
ServicebetaPythonDecision middleware for AI agents that uses local Laya to pick tools and judge results, falling back to an LLM when unsure.
ProjectbetaPythonSelf-improving multi-platform agent harness with an optional Jev companion for typed judgments and per-session model routing.
ServicebetaTypeScriptOmO/senpi agent plugin where Jev or Span-01 rates which skills and next tools fit the task and spots loops or completion.
Agent skillbetaTypeScript