Feynman — Research Agent Cheat Sheet

What Feynman is

Feynman is an open-source, terminal-first AI research agent (MIT licensed, companion-inc/feynman). Point it at a question and it searches papers (via alphaXiv), crawls the web, runs optional compute in Docker, and returns source-grounded briefs with citations — not unsupported essay prose.

Built on the Pi agent runtime. Docs: feynman.is/docs.

MyGearHut uses Feynman for competitive intel, technical due diligence, and literature sweeps where "chatGPT said so" is not enough.

Four subagents (automatic dispatch)

| Agent | Job |

|-------|-----|

| Researcher | Gather evidence from papers, web, repos, docs |

| Reviewer | Critique gaps, severity-grade weaknesses |

| Writer | Structure drafts from notes |

| Verifier | Check citations, URLs, dead links |

You do not manually route them — workflows invoke the right mix.

Install (macOS / Linux / Windows)

curl -fsSL https://feynman.is/install | bash

Then configure a model provider:

feynman setup

Supports Anthropic, OpenAI, Google, OpenRouter, and others per docs. Pick the model your budget and quality bar agree on.

First run — interactive REPL

feynman

Inside the REPL, slash commands trigger workflows:

/deepresearch transformer architectures for protein folding
/literature [topic]
/audit [claim you want verified]

Type /help for the full workflow list on your installed version.

One-shot from shell

feynman deepresearch "Compare RAG chunking strategies for legal PDFs — pros, cons, citations"

Outputs land under project outputs/ (path may vary — check docs). Every research claim should carry a retrievable source URL.

Local workbench (optional GUI)

For chat + artifacts + notebooks in one UI:

feynman serve

Opens the standalone science workbench — useful when you want provenance trails and file browsing, not just terminal scrollback.

Core workflows (operator view)

Deep research

Best for: strategic questions, market/tech surveys, lit reviews.

/deepresearch [precise question with scope and date preference]

Ask for scope: "peer-reviewed only", "post-2023", "implementation-focused".

Literature review

Best for: annotated bibliographies before you write.

/literature [topic keywords]

Expect structured summaries with paper links — still verify critical claims yourself.

Replication / experiments

Feynman can plan and run experiments in Docker or cloud GPU environments when configured. Treat this as Yellow zone — human must review methods and safety.

Requires Docker installed locally for container paths; cloud paths need provider tokens per docs.

Audit / verify

Feed a draft or claim set; Verifier stresses citations and flags unsourced statements.

Environment variables (common)

| Variable | Purpose |

|----------|---------|

| Provider API keys | Set during feynman setup |

| HF_TOKEN / HUGGINGFACE_HUB_TOKEN | Hugging Face Hub inspection tools |

| Docker socket | Local experiment execution |

Never commit keys; use shell profile or secret manager.

Prompt discipline for research agents

Feynman reduces hallucination but does not eliminate it. Front-load:

Scope: [narrow question]
Sources: peer-reviewed preferred; include grey literature if <5 papers
Output: executive summary + table of sources + open questions
Exclude: marketing pages unless noting bias
Date floor: [year]

Quality checklist before you ship a brief

• Every material claim has a URL or DOI
• You spot-checked top three sources manually
• Conflicting studies noted, not averaged away
• Methodology limits stated (sample size, benchmark choice)
• No confidential client data in prompts

Feynman vs generic chat

| | ChatGPT / Claude chat | Feynman |

|---|----------------------|---------|

| Citations | Inconsistent | Designed source-grounded |

| Paper search | Generic browse | alphaXiv integration |

| Workflow | Ad hoc | Slash research pipelines |

| Compute | Limited | Docker / GPU paths |

Use chat for drafts; use Feynman when citations and method matter.

MyGearHut research loop

1. Frame question in one paragraph (ICP for intel)
2. Run /deepresearch with date + source constraints
3. Human pass — read top sources, not just summary
4. Extract actions for your Notion/CRM — Feynman does not replace judgment
5. Archive output path in project wiki

---

Install once, cite everything. Browse more free guides at MyGearHut library.