The Workbench · Working Knowledge

Before you compare

Three separations that matter.

Most product debates collapse layers. Keep these distinctions straight and the cards below become readable.

A model is not a tool

GPT, Claude, Gemini, and Llama are model families. They do not, by themselves, open your repo, run tests, or file a pull request. A tool or harness wraps a model with context, actions, execution, memory, permissions, and verification. Comparing “Claude vs Cursor” is often comparing different layers.

A chat app is not an agent

A chat window answers questions. An agent can take multi-step actions: read files, run commands, call APIs, and loop until a task completes. The gap is not intelligence — it is what the system is allowed to do and where it runs.

Compare harness layers, not hype

The useful questions: What context can it see? What tools can it call? Where does execution happen? How are project instructions stored? What must you approve? How does it verify its work? The cards below use those layers and link to the first-party material checked for each product.

The comparison model

Six layers around the model.

The model is the engine. The harness is the cockpit. The workbench decides what the AI can see, touch, remember, and verify.

Model

GPT · Claude · Gemini · Llama · Qwen · DeepSeek · Hermes

Context

Files · repo index · long context · RAG · web search · MCP · screenshots

Actions

Shell · browser · git · APIs · databases · messaging · project systems

Execution

Local terminal · IDE · cloud sandbox · Docker · SSH · CI

Rules & memory

Project instructions · saved context · session history · external state

Control

Permissions · approvals · sandboxes · tests · diffs · logs · rollback

The products

Compare by working style.

Filters change only this list. Open a card for the full context, permission, verification, and source record.

Terminal pair-programming agent

Aider

A Git-centered terminal coding assistant that builds a concise repository map, edits selected files, commits changes, and can automatically lint and test them.

Verified July 13, 2026 · Open card →Multi-surface coding agent

Claude Code

An Anthropic coding agent that works from the terminal and connected development surfaces, with shell, file, Git, MCP, and configurable permission workflows.

Verified July 13, 2026 · Open card →Coding agent across CLI, IDE, app, and cloud

OpenAI Codex

OpenAI's connected coding-agent family for local collaboration and cloud delegation, with sandboxing, approvals, repository instructions, and evidence from files and command output.

Verified July 13, 2026 · Open card →IDE-native agent harness

Cursor

An AI code editor centered on interactive repository search, multi-file edits, terminal tools, checkpoints, reusable rules, and in-editor diff review.

Verified July 13, 2026 · Open card →Google terminal and IDE coding agents

Gemini CLI / Code Assist

Google's open-source Gemini CLI and Code Assist agent mode share file, shell, search, MCP, planning, and approval patterns across terminal and supported IDEs.

Verified July 13, 2026 · Open card →Open-source personal and multi-channel agent

Hermes Agent

A Nous Research agent runtime with a terminal UI, messaging gateway, tools, skills, memory files, model-provider choice, scheduling, and configurable approvals.

Verified July 13, 2026 · Open card →A category of models and inference runtimes

Open-weight / local models

A do-it-yourself stack in which model weights and a chosen runtime—such as llama.cpp, Ollama, or LM Studio—run on hardware you control; capabilities depend on every component.

Verified July 13, 2026 · Open card →Self-hosted personal-agent gateway

OpenClaw

An open-source gateway that connects messaging channels to an always-available agent with sessions, tools, memory, plugins, and multi-agent routing.

Verified July 13, 2026 · Open card →Open-source software-development agent platform

OpenHands

A developer-agent platform with SDK and application surfaces, isolated runtimes, workspace abstractions, command and file tools, and local or remote deployment options.

Verified July 13, 2026 · Open card →Research, search, and grounded-answer layer

Perplexity

A research product and developer platform built around current web search, extracted source material, citations, and generated synthesis rather than repository execution.

Verified July 13, 2026 · Open card →

Product examples are illustrations, not endorsements. Capabilities change — check sources before you rely on them.

Fit matrices

Match the job, not the brand.

Four common tasks with honest fit notes. Use these before you open a product card.

Fix a failing auth test

IDE agentBest when you are actively steering diffs in an editor.
Terminal coding agentBest when the task is command-heavy and repo-wide.
Cloud delegated agentAcceptable when you can review a PR later.
Research / searchPoor fit alone — use before coding if the failure is API-related.

Research which API changed last week

Research / searchBest before implementation — citations and freshness matter.
Terminal coding agentAcceptable after you know what to look for.
IDE agentPoor fit as the first step — not a research layer.
Personal agentPoor fit — not built for sourced technical lookup.

Turn a screenshot into a working UI

IDE agentBest when you iterate visually in the editor.
Terminal coding agentAcceptable if browser or image tools are in the harness.
Cloud delegated agentAcceptable for a first pass you review manually.
Research / searchPoor fit — this is implementation, not lookup.

Run a weekly dependency audit

Cloud delegated agentBest when the task can run on a schedule and return a report or PR.
Personal agentBest for reminders and routing across channels.
Open-source / self-hosted stackBest when you need control over what runs and where.
IDE agentPoor fit for unattended weekly runs.

Decision guide

Start from the work.

I want to actively code in an editor.

Cursor, Windsurf, and similar IDE-native harnesses.

I want a terminal agent to inspect, edit, and run my repo.

Claude Code, Codex CLI, Aider.

I want to delegate a task and review a PR later.

Codex web/cloud, Claude Code background modes, OpenHands cloud or self-hosted.

I want current research with citations.

Perplexity, Gemini search-grounded workflows, ChatGPT Deep Research.

I want a personal assistant across messaging apps.

Hermes Agent, OpenClaw.

I want self-hosted agent infrastructure.

OpenHands, OpenClaw.

I want lightweight terminal pair programming.

Aider, Codex CLI.

I want local, private, or open-weight model control.

Ollama, LM Studio, llama.cpp, vLLM, SGLang.

Product examples are illustrations, not endorsements. Capabilities change — check sources before you rely on them.

If you already know a tool

Translate between harnesses.

Short bridges for readers arriving with one product already in their hands.

If you know Cursor, what does Claude Code add?

Cursor is optimized around editor-native search, diffs, checkpoints, rules, and terminal tools. Claude Code begins from an agent-oriented terminal workflow and connects to additional development surfaces. The practical shift is from steering inside an editor to choosing where an agent executes and how its tools are approved.

If you know Claude Code, what does Codex add?

Both are coding-agent harnesses with overlapping local workflows. Codex also emphasizes connected cloud delegation. Compare the selected execution environment, sandbox, approval policy, tool integrations, and repository-instruction format for the task in front of you.

If you know Codex, what does Cursor add?

Codex spans CLI, IDE, and cloud delegation. Cursor optimizes for in-editor pair programming — semantic search, diff review, and tight feedback loops while you type. Use Cursor when the task is interactive editing; use Codex-style delegation when the task can run separately.

If you know Perplexity, what do coding agents add?

Perplexity is a research and search layer — web grounding, citations, APIs. Coding agents add repo context, file edits, shell execution, and verification loops. Use research tools before you build; use agents when the work touches your codebase.

If you know closed tools, what does open-source / local add?

Open and self-hosted stacks trade convenience for control: model swapping, private inference, custom sandboxes, and your own security boundary. You own more of the permission ladder — which is powerful and more work.

If you know chat assistants, what does an agent runtime add?

Chat answers in a thread. An agent runtime adds tool calls, persistent project instructions, execution environments, and loops that continue until a task completes or hits a guardrail. The model may be the same; the harness is not.