Reasoning as a Graph.

Advanced AI orchestration via Graph of Thoughts (GoT). Turn objectives into explicit plans. Execute, observe, and persist every step.

How it works

CORE FEATURES

Graph of Operations

Execute modular plans using Generate, Aggregate, Refine, and Score nodes. Static GoO structures define the reasoning logic.

Reasoning State (GRS)

Dynamic Graph Reasoning State persists every "thought" and dependency. Inspect and resume runs from any historical point.

Parallel Execution

Leverage max_parallel to execute independent reasoning branches concurrently, maximizing volume while minimizing latency.

Model Integration

Crossing the model boundary through isolated subprocesses. Native support for input and output contracts via Codex.

Sandbox Isolation

Automatic path mapping and sandbox hints ensure file operations remain within task working directories.

Audit Trail

Every control-plane boundary emits an event to events.jsonl. Mirrored to SQLite for advanced observability.

Benchmark

Based on ETH Zurich research data. GoT represents the global optimum for complex reasoning tasks.

Reasoning: Latency vs Volume

THEORETICAL

GoT achieves full volume (N) with logarithmic latency, outperforming CoT (sequential) and ToT (limited volume).

Quality vs Cost (Sorting 128)

EXPERIMENTAL
+62%
Quality Improvement (vs ToT)
>31%
Cost Reduction (vs ToT)

Get Started

zsh — 80x24
➜ npm install -g github:osvaldoandrade/neuron

remote: Enumerating objects: 42, done.

remote: Total 42 (delta 28), reused 0 (delta 0), pack-reused 0

added 1 package in 1.2s

➜ neuron

Neuron is a long-running task reasoner and orchestrator driven by Graph of Thoughts.

Usage:

neuron [flags]

neuron [command]

Available Commands:

completion Generate the autocompletion script goo Graph of Operations helpers graph Graph operations grs Graph Reasoning State helpers help Help about any command init Initialize workspace for Neuron logs Show task logs plan Generate or refine a graph without executing run Plan and execute until completion shell Start interactive shell task Task operations theme Theme helpers version Show version

Flags:

--config string Config file path --dry-run Plan only, do not execute -h, --help help for neuron --json JSON output --log-level string debug|info|warn|error --max-iterations int Max graph refinement iterations --model string Model override (ex: gpt-5-codex-mini) --no-color Disable color output --strategy string Reasoning strategy: cot|got --theme string UI theme: default|ocean|amber --timeout duration Timeout per step (e.g. 10m) --workdir string Base working directory

Use "neuron [command] --help" for more information about a command.

➜ neuron run "Prove P=NP :-)"
Reasoning Engine Active
Strategy: GoT / Runtime: GoO

Technical FAQ

What is a Graph of Operations (GoO)?
The GoO is a static dependency graph of operation types (Generate, Score, KeepBest, Aggregate, Refine, Validate). It serves as the blueprint for how Neuron transforms thoughts into solutions.
How does the GRS differ from a classic work graph?
The Graph Reasoning State (GRS) is dynamic. While the GoO defines the *how*, the GRS captures the *what*—the actual thoughts generated, their scores, and their relationships (refined_from, aggregated_from, etc.).
Can I pause and resume a GoO run?
Yes. Since Neuron persists the GRS and operation status after each step, the neuron task resume command will re-enter the scheduler and continue exactly where the reasoning chain left off.