WHEELER

Reliable, trustworthy, trackable AI workflows for science.

Runs 100% locally. No API keys, no cloud services. Your data never leaves your machine.

Named after great physicist John Archibald Wheeler, Niels Bohr's longtime collaborator. Wheeler and Bohr worked by talking. Bohr would pace, thinking out loud. Wheeler would push back, sharpen the question, sketch the math. The best ideas emerged from the conversation, not from either person alone. That's the model here.

Quick Start

uvx wheeler init my-research-project
cd my-research-project && claude
/wh:start

That's it. The first command scaffolds the project (.plans/, .wheeler/, wheeler.yaml, .mcp.json) and installs slash commands and agents to ~/.claude/ unless the wh plugin is already providing them. The second drops you into Claude Code with Wheeler's MCP servers wired up. The third routes you to the right /wh:* command for what you want to do.

For long-lived use install Wheeler globally (faster startup, stable paths in .mcp.json):

uv tool install wheeler
wheeler init my-research-project

Run wheeler doctor any time to verify your setup: Python version, deps, Claude Code, Neo4j connectivity and whether the connection is TLS, which project-isolation model is in force, and whether a legacy install is shadowing the plugin.

Prerequisites: Python 3.11+, uv, Claude Code (Max subscription), and Neo4j Desktop (free). New to all this? Walk through the Getting Started Guide.

Installing the acts

Wheeler's /wh:* acts ship as a plugin. In Claude Code:

/plugin marketplace add maxwellsdm1867/wheeler
/plugin install wh@wheeler

Codex is supported as a host too, with the same two steps:

codex plugin marketplace add maxwellsdm1867/wheeler
codex plugin add wh@wheeler

The plugin updates itself and works in every project. It ships from v0.14.0. To try it from a clone instead, point your host at the checkout: claude --plugin-dir <path-to-clone>, or pass that path to codex plugin marketplace add (it takes a local path as well as owner/repo).

wheeler install is the LEGACY path: it copies the act files into ~/.claude/commands/wh/. Those files SHADOW the plugin. Claude Code resolves /wh:plan to a file in ~/.claude/commands/wh/ before it looks at the plugin's skills, with no error and no warning, so a machine with both keeps running stale local copies of every act and never sees a plugin update.

If you installed Wheeler before the plugin existed, switch over with:

wheeler migrate-to-plugin

It lists exactly what it will remove (act files, agents, hooks, the statusLine entry, the wheeler_* MCP registrations), asks once, then prints the commands above. It is idempotent and safe to run when there is nothing to migrate. wheeler install now refuses to run while the plugin is present rather than silently shadowing it, and wheeler doctor reports the collision if you ever end up in it.

Neo4j credentials

wheeler login stores them in the OS keychain instead of a shell profile. Easiest route is Aura's own credentials file:

wheeler login --aura-file neo4j-credentials.txt   # or bare `wheeler login` to type the fields
wheeler login --status                            # which of env, keychain, wheeler.yaml, or default supplies each setting

The credential is validated by connecting before it is stored, and the password is never written to a file or echoed. Environment variables still win over the keychain, so existing setups keep working. See the Getting Started Guide for the full walkthrough.

From source

git clone https://github.com/maxwellsdm1867/wheeler.git
cd wheeler
uv sync --extra dev              # editable install + tests + ruff + mypy + build
uv run wheeler init ~/my-research-project

bin/setup.sh is still around for the full bootstrap (Neo4j in Docker, schema init, git hooks, zsh completions).

Why Wheeler

Science requires reproducibility. As AI gets embedded in research workflows, the gap between "AI helped me" and "here's the auditable chain of how this result was produced" becomes a credibility problem.

Wheeler is built on four pillars:

Traceable results. When Wheeler creates a finding, it automatically records what script ran, what data it consumed, what papers informed the approach, and when it happened. One tool call builds the full provenance chain. The agent focuses on science; infrastructure handles bookkeeping.

Change propagation. When a script changes or data is updated, Wheeler flags every downstream finding as stale and reduces its stability score. You always know what to trust and what needs re-verification.

Context management. All components read from and write to the same graph, so a finding from data analysis immediately informs subsequent literature searches, experimental design, and manuscript preparation. Information is progressively disclosed and retrieved only when relevant.

Executable research artifact. The knowledge graph moves beyond the static PDF. It is an executable map of discovery: any scientist can inherit the full experimental context of a project, explore how results connect, and build directly on top of prior work.

The Workflow

Wheeler gives you a fluid cycle, not a rigid pipeline. Enter at any point, skip stages, repeat them.

 TOGETHER         you + wheeler, thinking out loud
 discuss  plan  chat  pair  write  note  ask
                         |
                         v  remaining work is grinding
 HANDOFF          propose independent tasks
 handoff          you approve, modify, or keep talking
                         |
                         v
 INDEPENDENT      wheeler works alone
 wh queue "..."   logged, stops at decision points
                         |
                         v
 RECONVENE        results + flags + surprises
 reconvene        back to TOGETHER

Every plan and execution renders a self-contained visual brief: the question and sub-questions, figure mockups (pre-registered sketches) paired with the real result figures, a pipeline flow chart, and the data sources. /wh:discuss reads that brief to interpret the results with you like a colleague, referencing figures by number and running quick checks against the data to strengthen or disprove a point.

A typical session

The flow we design for, end to end:

1. /wh:discuss — talk through the question until it is sharp. Wheeler asks like a colleague, grounds the conversation in what the graph already knows, and locks the decisions.
2. /wh:plan — Wheeler structures the investigation into waves of tasks and, before any data is touched, pre-registers the figures: what each one plots and how competing hypotheses would look different in it. On approval it renders a visual brief (question, mockups, pipeline, data sources) so you react to a picture, not prose. Seeing the mockup often sends one more round of sharpening back into the plan.
3. /wh:execute — Wheeler runs the WHEELER-assigned tasks, logs findings with full provenance, then regenerates the brief as a report: each pre-registered mockup now sits beside its real result figure, success criteria are marked, and result tables tuck into dropdowns.
4. /wh:discuss (again, on the results) — hand Wheeler the brief and interpret together: what holds, what is fragile, what the next question is. Wheeler references figures by number, pulls related findings from the graph, and can run a quick check against the data to settle a contested point, registering whatever you endorse back into the graph.
5. /wh:write drafts from the endorsed findings with strict citations, or /wh:plan opens the follow-up investigation. /wh:close sweeps the session into a synthesis.

You can enter at any step, skip stages, or loop steps 2 to 4 as the work demands.

Commands

Headless mode

Wheeler can run tasks without you present:

wh queue "search for papers on SRM models"   # sonnet, 10 turns, logged
wh quick "check graph status"                 # haiku, 3 turns, fast
wh dream                                      # graph consolidation

The wh launcher is a bash script in bin/wh that ships only with the source tree, not the PyPI wheel. To enable it after a uv tool install, clone the repo and symlink it: sudo ln -sf $PWD/bin/wh /usr/local/bin/wh. A native wheeler queue / quick / dream is on the roadmap.

Wheeler never does your thinking. Every task gets tagged: SCIENTIST (judgment calls), WHEELER (grinding), or PAIR (collaborative). Decision points are flagged as checkpoints, not guessed at.

How It Works

Provenance-completing tool calls

The core primitive: one tool call creates a finding AND its full W3C PROV-DM provenance chain. You never write this directly; slash commands handle it. But under the hood, this is what happens:

add_finding(
    description="Midget and parasol cells have similar clusters of fitted SRM parameters",
    confidence=0.85,
    execution_kind="script",                    # auto-creates Execution activity
    used_entities="D-abc123,S-def456",          # auto-links inputs
)

Wheeler internally creates the Finding, an Execution activity node, links inputs (Dataset, Script) via USED, links the output via WAS_GENERATED_BY, sets a stability score, and dual-writes to Neo4j and JSON. The provenance chain is always complete because the agent never had to remember to create it.

Stability and invalidation

Every entity carries a stability score (0.0-1.0) encoding epistemic trust: primary data = 1.0, published papers = 0.9, validated scripts = 0.7, LLM-generated findings = 0.3. When an upstream entity changes, stability decays downstream: new = source * (0.8 ^ hops). Changed scripts propagate stale flags through the entire dependency chain.

The knowledge graph

The graph is an index over files, not a document store. Each node stores an ID, type, tier, title, path, and timestamps. Full content lives in knowledge/{id}.json. Human-browsable rendering lives in synthesis/{id}.md (Obsidian-compatible with YAML frontmatter and [[backlinks]]). When you need connections, ask the graph. When you need content, read the file.

11 entity types: Finding, Hypothesis, OpenQuestion, Dataset, Paper, Script, Execution, Document, ResearchNote, Plan, Ledger.

14 relationship types: 6 W3C PROV standard (USED, WAS_GENERATED_BY, WAS_DERIVED_FROM, WAS_INFORMED_BY, WAS_ATTRIBUTED_TO, WAS_ASSOCIATED_WITH) + 8 Wheeler semantic (SUPPORTS, CONTRADICTS, CITES, APPEARS_IN, RELEVANT_TO, AROSE_FROM, DEPENDS_ON, CONTAINS).

51 MCP tools across 4 servers (core, queries, mutations, ops).

See ARCHITECTURE.md for the complete technical spec: module dependency map, PROV schema, MCP tool listing, hardening patterns, design decisions.

Service integrations

External research tools land in the graph as provenance-tracked nodes. The model is a sandwich: an act reads graph context and shapes the request, the tool's own CLI runs (owning its auth and retries), and one deterministic Python ingest writes the result back through the triple-write. Every call is one Execution whose status is truthful: a failed or incomplete job is recorded as failed with no fabricated outputs (the external-call failsafe), never masquerading as a clean run. Four Ai2 Asta services ship today (Paper Finder, Semantic Scholar, Theorizer, Literature Reports), routed by /wh:asta.

Adding a new service is its own loop: the wheeler-service-creator skill scaffolds the adapter (registry contract, ingest, act, and test) with the failsafe baked in, then a bundled auditor checks data-safety, provenance, and conventions before it lands. See ARCHITECTURE.md "Service Integrations".

What's New

Architecture

Claude Code (interactive)
    ├── /wh:* slash commands (.claude/commands/wh/*.md)
    │       ├── /wh:start: intent router (invokes other commands)
    │       ├── YAML frontmatter: tool restrictions per mode
    │       └── System prompt: workflow + provenance protocol
    │
    ├── MCP Servers (54 tools)
    │       ├── wheeler_core (14): health, status, context, search, cypher
    │       ├── wheeler_query (11): read-only query_* tools
    │       ├── wheeler_mutations (19): add_*, link, delete, update, merge
    │       └── wheeler_ops (10): staleness, citations, consistency
    │
bin/wh (headless)
    └── claude -p with structured logging → .logs/*.json

wheeler/
├── models.py                # Pydantic v2: 11 node types, prefix mappings
├── config.py                # YAML loader, Pydantic config models
├── provenance.py            # Stability scoring, invalidation propagation
├── consistency.py           # Cross-layer drift detection and repair
├── mcp_core.py              # Split server: health, context, search (12)
├── mcp_query.py             # Split server: query_* read-only (11)
├── mcp_mutations.py         # Split server: add_*, link, delete, update (18)
├── mcp_ops.py               # Split server: staleness, citations (10)
├── mcp_shared.py            # Shared: trace IDs, decorators, config
├── knowledge/               # File I/O: read, write, list, render, migrate
├── graph/                   # Neo4j backend, circuit breaker, schema, context
├── search/                  # Embeddings, RRF fusion, graph-expanded search
├── validation/              # Citation validation, ledger quality metrics
├── tools/graph_tools/       # Provenance-completing mutations + queries
└── workspace.py             # Project file scanner

tests/                        # 3160 tests
docs/                         # Getting started, architecture, project spec

Contributing

Bug reports: Use /wh:dev-feedback from inside a session to file structured issues, or report at GitHub Issues.

Tests: python -m pytest tests/ -v (3160 tests). E2E tests require a running Neo4j: python -m pytest tests/e2e/ -v.

Architecture: See ARCHITECTURE.md for the full technical spec (module dependency map, PROV schema, MCP tool listing, hardening patterns).

Project docs:

• Mission — four pillars, target audience, design north star
• Tech stack — components, infrastructure patterns, current gaps
• Roadmap — shipped versions, v0.9.0 phases, v1.0 criteria
• Getting started — install walkthrough with Neo4j Desktop
• Project spec — original design specification

Citation

If you use Wheeler in your research, please cite it:

@software{hong_wheeler_2026,
  author    = {Hong, Arthur and Rieke, Fred},
  title     = {{Wheeler: Reliable, trustworthy, trackable AI workflows for science}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20498885},
  url       = {https://doi.org/10.5281/zenodo.20498885}
}

Integrations

Wheeler integrates with external research services so their output lands in the knowledge graph as provenance-tracked nodes, and so those services can act on Wheeler's own work and context. The first integration is AllenAI Asta: Wheeler ships tools (adapters) for four Asta services, Paper Finder, Semantic Scholar, Theorizer, and Literature Reports, routed by /wh:asta. Each call reads the current graph to shape the request, runs the Asta service, and writes the result back with full provenance (what it USED, what it WAS_GENERATED_BY, and how the new results connect to the existing graph). A failed call is recorded as failed rather than silently lost.

The integration layer is provider-agnostic and growing. Adding a new external tool is its own workflow: the wheeler-service-creator skill scaffolds the adapter, bakes in the provenance and failsafe wiring, and audits it before it lands. See ARCHITECTURE.md "Service Integrations" for the design, and the roadmap for where this is headed.

Acknowledgments

Wheeler's Asta integration shells out to the Asta toolkit from the Allen Institute for AI (Ai2). The Paper Finder, Semantic Scholar, Theorizer, and Literature Reports services are Ai2's work (asta.allen.ai); Wheeler does not vendor or reimplement them, it invokes the upstream asta CLI and marshals the results into the knowledge graph with provenance. Credit and thanks to the Ai2 Asta team.

Wheeler's equation discovery is a plug-in for LLM-SR, not a method of Wheeler's own. /wh:llmsr-discover is an adapter: it drives the LLM-SR pipeline from Claude Code and records the result with provenance. LLM-SR (MIT) is the work of Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K. Reddy, published at ICLR 2025 as "LLM-SR: Scientific Equation Discovery via Programming with Large Language Models", and it builds in turn on FunSearch (Apache-2.0) from Google DeepMind, published in Nature as "Mathematical discoveries from program search with large language models".

The adapter exists because Wheeler runs on a Max subscription with no API keys, so upstream's sampler and its orchestration loop cannot be used as shipped. wheeler/integrations/llmsr/vendor/ carries the six modules the driver needs, adapted from their pipeline; the substituted modules are the sampler and the loop. Every other change to their code is environmental (package-relative imports, Python 3.12 AST renames, dropping the absl and torch dependencies, a macOS fork context, and a numpy softmax so scipy stays optional). The search algorithm, the island model, and the program-manipulation logic are theirs, untouched. One further piece is substituted rather than changed, and only by default: the SCORING seam. A spec declares an @evaluate.run that fits a candidate's constants and returns its score, and by default the driver does not call it, scoring instead through Wheeler's own fit/metric seam so the metric is pluggable, the fitted constants are recoverable, and each group can refit its own. A run created with wheeler llmsr init --use-spec-evaluate calls the spec's own @evaluate.run instead, in upstream's data shape and reading upstream's bare-float return, so a spec that trains its own model inside evaluate runs unmodified. The door is a declared flag, never inferred from the spec text. Their scoring code is here unaltered; what changes is only whether the driver takes that path.

If you publish results from /wh:llmsr-discover, cite LLM-SR (and FunSearch where appropriate), not Wheeler. For the complete and actively maintained implementation, the benchmark problems, and the authors' own documentation, go upstream. Full per-file attribution, both licenses, and BibTeX entries are in wheeler/integrations/llmsr/vendor/NOTICE.md. Credit and thanks to the LLM-SR authors and to the FunSearch team.

License

MIT