What AI tools are we using for development in 2025?
8 min read
Sep 20, 2025
AI
Development Tools
Claude
Copilot
Cursor
Cline
Our day-to-day stack blends Claude Code (with Agent-OS specs), OpenAI Codex, Cursor, GitHub Copilot, and Cline. Together they cover end-to-end tasks: planning, project management, coding, refactoring, PR reviews, and secure fixes.
AI-assisted coding has moved from "nice to have" to "core toolchain." Below — the stack we actually use day-to-day in 2025, where each tool fits, what it's great at, and quick recipes you can copy.
TL;DR (our stack at a glance)
- • Claude Code + Agent OS — Deep refactors and repo-wide changes with predictable plans and standards.
- • Codex (OpenAI) — Agentic runs in a safe sandbox (Web) or locally (CLI) for task-oriented builds and test loops.
- • Cursor — The AI IDE for everyday editing, multi-file rewrites, and a fast "agent mode" with you in the loop.
- • GitHub Copilot — The always-on co-pilot for autocompletion, chat across IDEs, and emerging "agent" workflows tied to your repos.
- • Cline — Open-source coding agent in your IDE that can execute commands and use MCP tools, with client-side/BYOK architecture.
Claude Code + Agent OS: plan-driven repo surgery
Why we use it: When the task is bigger than a single file — onboarding to a large repo, triaging issues, or refactoring subsystems — Claude Code maps the codebase and reasons through multi-step changes surprisingly well. Pairing it with Agent OS gives us structure: checklists, standards, and repeatable workflows so results match our way of shipping.
Where it shines:
- – Codebase mapping & explanations without hand-curating context.
- – Agentic refactors guided by best-practice prompts (Anthropic's own playbook helps).
- – Org standards baked in via Agent OS specs and checklists.
Example workflow:
- 1. In Agent OS, define a "service extraction" playbook (inputs: target module, interfaces, testing rules).
- 2. Run Claude Code with that plan: generate patch, run tests, produce PR description.
- 3. Review diffs + commit.
When not to use: Tiny fixes; you'll move faster in Cursor or Copilot.
Note: Anthropic's 2025 model updates vastly expanded context and long-running agent workflows, which is why Claude Code handles big repos so well.
Codex (OpenAI): agentic runs in a sandbox or on your machine
What it is: In 2025, OpenAI reintroduced Codex as an agentic coding tool with two faces:
- • Codex Web — runs in a cloud VM, executes code, navigates directories, and reports every step.
- • Codex CLI — a local agent you install to work against your filesystem and tooling.
Why we use it: For task-scoped builds — e.g., "add CSV import with validation, create tests, run them, and output a patch" — Codex's environment and execution loop are convenient and auditable.
Cursor: the AI IDE you actually live in
Why we use it: Cursor is our daily driver for editing. It understands the repo, applies multi-line edits, and its agent mode completes end-to-end changes while keeping you in the loop (hit Ctrl+I). Great for "make this change across files" speed runs.
What we lean on:
- – Smart rewrites & multi-file diffs during refactors.
- – Ask the codebase for where a pattern lives and jump to edits.
- – Agent mode for guided tasks, still review-first.
Mini-recipe:
"Replace our custom debounce with lodash/debounce in all React components, keep behavior, add a single unit test per component, and update imports."
GitHub Copilot: the ambient assistant tied to your repos
Why we use it: Copilot is the "always-on" layer — autocomplete, inline chat, and chat in every major IDE and even Windows Terminal. It's the fastest path to fixing small things, exploring APIs, or explaining unfamiliar code right where you're typing.
What's new & useful:
- – Copilot Chat everywhere (GitHub.com, IDEs, mobile, terminals).
- – Agent/coding mode (preview) is rolling out; worth watching as it matures, especially inside PR and code-review flows.
Everyday prompts:
- – "Explain this diff like I'm new to the project."
- – "Write table-driven tests for these edge cases and open a draft PR."
When not to use: Long, multi-step tasks that need explicit planning — hand those to Claude Code/Agent OS or Codex.
Cline: open-source, transparent agent inside your IDE
What it is: Cline is an autonomous coding agent as a VS Code/IDE extension. It can create/edit files, execute terminal commands, browse (with your permission), and orchestrate tools through Model Context Protocol (MCP). You pick the model (BYOK), and it runs client-side — which is a big deal for security/compliance.
Why we use it:
- – Plan/Act modes for controlled autonomy.
- – Zero-trust design: your code doesn't leave your machine; enterprise-ready with SOC 2 compliance capabilities.
- – Great for enterprises wanting transparency and cost control (teams/BYOK).
Try this:
"Plan a migration from axios to fetch with a thin wrapper. Show the plan first. Then apply changes in /api, update tests, run pnpm test, and stop on any failure with a summary."
How we choose the right tool for the job
| Scenario | Best fit | Why |
| Repo-wide refactor with standards & sign-offs | Claude Code + Agent OS | Plans + large-context reasoning with repeatable playbooks. |
| "Build X, run tests, give me a patch" | Codex (Web/CLI) | Sandboxed execution and clear, auditable steps. |
| Daily edits, multi-file rewrites, quick tasks | Cursor | AI IDE built for fast editing and agent-assisted changes. |
| Inline help, autocomplete, PR summaries | GitHub Copilot | Ambient assistant tied to GitHub + IDEs. |
| Autonomous agent with client-side control | Cline | BYOK, terminal execution, MCP tools, enterprise-friendly. |
Practical tips that saved us time
- • Show your work: Prefer agents that explain plans and commands (Codex Web, Cline Plan Mode, Claude Code with verbose traces). This makes reviews fast and safe.
- • Codify standards once: Use Agent OS to encode naming rules, testing expectations, and PR templates; reuse across repos.
- • Use the right context window: Large-context models reduce the "chunking tax" for big changes. If the task spans many modules, pick Claude 4/Sonnet 4 era models via Claude Code.
- • Keep humans in the loop: Cursor's agent and Copilot's chat are fastest with tight prompts + quick reviews.
- • Security first: For sensitive code, favor client-side/BYOK setups (Cline; Codex CLI) and limit scopes/tokens.
Final take
No single tool wins every scenario. Our rule of thumb:
- • Think: standards & plans → Claude Code + Agent OS
- • Build & test in a box (or locally): Codex
- • Edit fast: Cursor
- • Stay productive everywhere: GitHub Copilot
- • Go autonomous with control: Cline
Used together, they feel less like "magic" and more like a reliable production line for software — exactly what you want in 2025.
Useful links
AI tools:
Ready to get started?
Let's discuss how AI can transform your business today