ApexCode AI Assistant
Most AI coding assistants lock you into one provider and act on your code before you've seen what they're planning. ApexCode does neither. Bring your own API key from any of 7 supported providers, and every mutating action waits for your approval first — right inside Visual Studio.
Why ApexCode
- No subscription, Bring-Your-Own-Key. You only pay your provider's token cost — no markup, no platform fee. Plug in a key from OpenAI, Azure OpenAI, OpenRouter, DeepSeek, Gemini, NVIDIA, or Vercel and you're running.
- Plan before it acts. Switch into Plan Mode and ApexCode explores your workspace read-only, proposes a plan, and waits for approval before any file is touched. Tool access itself is restricted, not just prompted around.
- Built for real codebases, not toy files. Large files are read in paginated line ranges, cached by last-write-time, and older conversation history is automatically evicted to stay within your provider's token limit — no context blowups on real-world projects.
- Talks to your database, safely. Connect to SQL Server to inspect schemas and run queries. Writes and stored procedures are blocked outright — only validated, read-only SELECT/CTE queries run, capped at 100 rows.
- Deep IDE awareness. Uses Roslyn to read diagnostics, validate builds, and apply fixes directly to your syntax tree — not paste-and-hope suggestions.
- Checkpoints, not regret. Every AI turn that uses tools takes an automatic workspace snapshot first. Review a diff, revert with one click.
- Custom personas. Define roles like "Angular Developer" or "SQL Reviewer" that shape how the assistant behaves, injected straight into its context.
Core Features
| Feature |
What it does |
| Multi-provider chat (BYOK) |
Connect any of 7 LLM providers with your own API key |
| Plan / Act modes |
Read-only exploration and approval required before any mutating action |
| Paginated file reading |
Large-file-safe context handling, cached and capped |
| SQL Server integration |
Read-only schema inspection and query execution, capped at 100 rows |
| Roslyn-powered fixes |
Compile validation and AI fixes applied to the syntax tree |
| Checkpoints & diffs |
Snapshot and revert any AI-driven change |
| Custom personas |
Tailor the assistant's behavior per project or task |
| Token budget management |
Automatic history eviction to fit provider context limits |
Requirements
- Visual Studio 2022 (or later, where supported)
- An API key from a supported LLM provider (OpenAI, Azure OpenAI, OpenRouter, DeepSeek, Gemini, NVIDIA, or Vercel)
- SQL Server connectivity is optional and only required for the database features
One flag: I left the version requirement as "Visual Studio 2022 (or later, where supported)" instead of naming VS 2026 outright — worth replacing with the exact version once you've confirmed it's actually tested there, since an unverified compatibility claim is an easy thing for a technical reader to catch and hold against the listing.
| |
