Feynman vs Kit: Features, Pricing & Which Is Better (2026)

A side-by-side comparison of Feynman and Kit — features, pricing, and ideal use cases — to help you decide which AI tool fits your workflow.

Feynman

Companion

Free

Open-source AI research agent that reads papers, ranks literature, drafts research and plans experiments from the terminal or a local workbench.

Cited Research Briefs: Asking a research question returns a synthesized brief where each claim is tied to the paper or web source it came from, rather than an unsourced summary.
PaperRank Scoring: Ranks papers on a topic with transparent evidence for citations, methodology, reproducibility and provenance so reading order is a decision you can inspect.
Paper Access Resolver: Resolves a single DOI, arXiv ID, OpenAlex ID, PMID, PMCID or title against OpenAlex, arXiv/alphaXiv, DOI and Europe PMC, with optional full-text fetching.
Local Science Workbench: `feynman serve` opens a standalone app with projects, sessions, chat, notebooks, compute, artifact previews and provenance in one place.
Claim Auditing and Replication: Compares a paper's stated claims against what its code actually does, and generates replication plans with compute targets and gated experiment steps.
Local and Hosted Models: Works with hosted providers via OAuth or API key and with local runtimes including LM Studio, Ollama, vLLM and a LiteLLM proxy.
Skills-Only Install: The research skill library can be installed on its own into Claude, Codex or OpenCode projects without the terminal app or bundled runtime.
Science Artifacts: Reports, data files, spreadsheets, notebooks, LaTeX, chemistry sketches and genomes are browsable together with versions, lineage and execution logs.

Deciding What to Read: Ranking a fresh literature pile on a topic by reproducibility and methodology instead of citation count alone.
Writing a Literature Review: Producing a review that separates where the field agrees from where questions remain open, with citations attached.
Verifying a Paper's Claims: Auditing whether the results a paper reports are supported by the code and data it released.
Planning a Replication: Turning a published finding into a concrete replication plan with a compute target and staged experiment steps.
Running Deep Research Passes: Launching a multi-agent deep dive on a topic that synthesizes findings and verifies them before reporting.
Keeping Research Local: Running the whole pipeline against a local model so unpublished work and private data never leave the machine.
Adding Research Skills to a Coding Agent: Installing the skills bundle into an existing Claude or Codex project to get research workflows without a second app.

Speakeasy

Free

An open-source coding agent runtime that gives the model a single compose tool, cutting round trips and token use versus conventional harnesses.

Single Compose Tool: The model receives one tool whose argument is a Runlet program, so file reads, shell commands, edits, retries, and subagent calls all happen in a single round trip instead of one per action.
Concurrent Program Execution: Independent calls inside a compose program run concurrently, with data dependencies or after blocks used to force ordering when needed.
Reusable Subagents: A subagent is a value you can continue, fork, inspect, and close, and you can require its output to match a JSON schema.
Cross-Harness Orchestration: Claude Code, Codex, Cursor, or another Kit instance can be used as the subagent harness over ACP, so Kit coordinates tools it did not ship with.
Open Protocol Support: Implements ACP v1 and v2 over stdio, HTTP/SSE, and WebSocket, A2A v1 in both directions, plus MCP, Agent Skills, and Agent Plugin packages.
Crash-Safe Long Sessions: Each append-only JSONL transcript item is synced to disk before acceptance and crash-safe locks let sessions resume from the TUI, prompt, or any ACP client.
Automatic Context Compaction: Context is compacted automatically at 80% of the model's context window so long runs do not stall on overflow.
Flexible Model Access: Connects to ChatGPT subscriptions through native OAuth and to models via OpenRouter or the Speakeasy AI Control Plane.
Single Static Binary: Ships as one binary with a published container image, avoiding a runtime dependency chain on developer machines and CI.

Cost-Sensitive Agentic Coding: Cut token spend and wall-clock time on large refactors by collapsing many tool calls into one composed program.
Editor-Integrated Agents: Drive Kit from any ACP-compatible editor without writing a bespoke integration for each client.
Multi-Harness Pipelines: Orchestrate Claude Code, Codex, or Cursor as subagents from a single controlling program when different harnesses suit different steps.
Long-Running Migrations: Run multi-hour codebase migrations that survive crashes and resume from a durable transcript.
Structured Extraction from Code: Require subagents to return schema-validated JSON so results can be fed into downstream tooling rather than parsed from prose.
CI and Headless Automation: Run the same binary in containers over HTTP/SSE or WebSocket to fix failing tests or apply mechanical changes without a terminal session.