Most “AI tools for developers” lists fail the same way: they group everything by category, give each tool a vague paragraph, and don’t tell you which ones senior engineers actually keep installed three months later. This cheatsheet is organized by workflow — the moments in your day when you reach for an AI tool — and it’s honest about which tools survive contact with a real codebase.
It covers six workflows: writing code, reviewing code, navigating unfamiliar code, debugging, automating ops, and connecting your AI to real systems via MCP. Each section names the tool you should try first, the realistic alternative, and the trap to avoid.
1. Writing code: IDE assistants
The IDE assistant is the most consequential AI choice you’ll make. It sets the ceiling on how much leverage you get from everything else.
Start here: Cursor. In 2026 Cursor is still the default choice for most teams. The composer, the @-mention context system, and the agent mode are mature, and the editor lineage from VS Code means almost zero migration cost. Expect $20/month per seat for the Pro tier; the free tier is real but you’ll hit the cap inside a week of serious use.
Realistic alternative: Claude Code. Anthropic’s CLI runs Claude inside a terminal pane that already has your repository’s context. If you live in tmux or you want a single long-running agent that drives multi-file changes, it’s a better fit than a popup-in-an-editor model. It plays nicely with Cursor — many engineers run both.
Trap to avoid: blindly trusting one model. Cursor’s model picker exists for a reason. Use Claude Sonnet for refactors and architectural reasoning, GPT-5-class models for boilerplate and language-specific autocomplete, and a fast local model for trivial completions. Locking yourself to one model means you’ll feel the failure modes of that model more than its strengths.
2. Reviewing code: PR review automation
Human review still matters, but AI review is now table stakes for catching the boring stuff. The goal is not to replace your senior reviewer — it’s to make sure your senior reviewer never has to comment “you have a typo on line 412” again.
Start here: CodeRabbit or Greptile. Both wire up in five minutes to GitHub, post inline comments on every PR, and learn your conventions over time. Greptile leans heavier on repo-wide context (it’s very good at “you’re re-implementing a helper that already lives in lib/utils”); CodeRabbit leans heavier on individual diff quality.
Built-in option: Vercel Agent or GitHub Copilot Review. If you’re already paying for the parent platform, the first-party review tooling has gotten genuinely good and avoids another vendor relationship.
Trap to avoid: turning on auto-comments for every PR immediately. AI reviewers are loud by default. Start with a quieter configuration (high-severity issues only) and dial it up once the team trusts the signal. The fastest way to make engineers ignore AI feedback is to drown them in nits.
3. Navigating unfamiliar code: search and explanation
When you join a new team, when you onboard a junior, when you crack open a dependency you’ve never read — these are the moments AI navigation tools earn their price.
Start here: Sourcegraph Cody or your IDE’s @ codebase feature. Both let you ask “where is the auth flow defined?” or “explain how the rate limiter works” against a real repo. Sourcegraph’s indexing is industrial strength; the IDE option is faster to set up and free.
Underrated: a single “tour” prompt at onboarding. Spend 20 minutes asking your assistant to produce a one-page map of the codebase — directories, top-level modules, the data flow, the half-finished migration nobody mentions. New hires can do this themselves on day one and it cuts ramp time in half.
4. Debugging: log analysis, trace explanations, root-cause
Debugging is where AI assistance is most under-marketed and most useful. A model that can hold a stack trace, three log files, and a recent diff in context at once will outperform a sleep-deprived senior engineer at the “is this a regression?” question.
Pattern: paste the failing test, the diff, and the relevant logs into your IDE assistant in a single message. Then ask: “What changed that could explain this?” Most of the time the model finds the issue in under a minute. The 1M token context windows in 2026 mean you almost never have to summarize.
Production debugging: Sentry AI, Datadog Bits, or Honeycomb Query Assistant. When the bug only happens in prod, an AI sitting next to your traces is worth a senior SRE you don’t have to page.
5. Automating ops: terminal AI and DevOps assistants
The shell is where AI assistance feels most magical when it works and most expensive when it doesn’t.
Start here: Warp or Fig with AI on. A terminal that suggests the right git rebase --onto incantation is a daily quality-of-life upgrade. Both have free tiers that are honestly enough for solo use.
For deployments: Claude Code with permission scopes. Claude Code can drive a real deploy — but only if you configure permissions carefully. Run it in plan mode first, and never give it write access to production without explicit per-action approval until you trust its judgment for your domain.
Trap to avoid: AI-generated CI configs. The model will happily generate a 200-line workflow file. Half of it will be cargo-culted from a Stack Overflow answer from 2022. Read every line before committing.
6. Connecting AI to real systems: MCP servers
Model Context Protocol is the connective tissue of 2026 AI tooling. An MCP server gives your Claude / Cursor / Windsurf assistant a typed, capability-scoped channel into a database, an API, or a filesystem. We have a dedicated guide on this — MCP Servers Explained — but the developer cheatsheet view is short:
- Filesystem MCP — read/write files outside the editor pane. Install this first.
- Postgres or SQLite MCP — your assistant can now inspect your schema and run queries. Game-changing for data work.
- GitHub MCP — open issues, comment on PRs, create branches from inside a chat session.
- Playwright or Browserbase MCP — your assistant can drive a real browser for testing or scraping.
- A vendor MCP for whatever you ship — Stripe, Linear, Vercel, Cloudflare, Webflow, Atlassian, and dozens of others now ship official MCP servers. Skim the catalog at repos.skila.ai/servers and install one that matches your stack.
Trap to avoid: installing every MCP server you find. Each MCP server is another permission surface and another way for an over-eager agent to do something you regret. Install deliberately, scope tightly, and audit every few months.
The 2026 starter stack
If you’re building a stack from scratch today, this is the minimum-viable configuration that handles 90% of cases:
- Cursor as the primary editor ($20/month per seat)
- Claude Code as the terminal companion (usage-based)
- CodeRabbit for PR review ($15/month per seat)
- Filesystem + Postgres + GitHub MCP servers wired into both
- One observability AI — Sentry AI if you’re on Sentry, Datadog Bits if you’re there. Skip if not.
Roughly $35–$55 per developer per month, all-in. That’s cheap compared to the time it saves any engineer worth keeping.
What to skip
Three categories of tools sound good but don’t earn their keep for most teams in 2026:
Universal “AI for everything” suites. The all-in-one platforms still ship below the per-workflow tools on most benchmarks. You’re paying a premium for one bill, not better output.
Single-purpose “AI test generator” tools. Your IDE assistant already does this, and the purpose-built tools tend to generate the same kind of brittle, over-mocked tests that humans regret writing.
“AI documentation” tools that regenerate README files on every push. Your docs drift not because nobody writes them, but because nobody reads them. Generating more of them does not fix the problem.
Keep this guide current
The AI tooling landscape moves fast enough that any cheatsheet is partly stale within a quarter. We update this guide when something meaningful changes — a new model tier, a category leader getting unseated, a pricing model shift. For the most current daily-level coverage of what’s actually shipping, follow news.skila.ai; for the live directory of tools mentioned here, browse tools.skila.ai.