Amp Code

Amp Code

Agentic coding tool from Sourcegraph. Canonical site: https://ampcode.com/. Positions itself between cursor (GUI, inline edits) and claude-code (fully autonomous CLI). The standout feature is think-out-loud mode — the agent narrates its reasoning as it works, making the process more auditable.

Think-Out-Loud Mode

Most agents show you the result, not the reasoning. Amp externalizes the chain of thought as it runs: what it's reading, why it's making a particular edit, where it's uncertain. This makes it easier to catch mistakes before they compound, especially on large refactors.

Think-out-loud is triggered by using "think hard" in your prompt, which invokes extended thinking mode. Useful when you want autonomous operation but still want to supervise the reasoning, not just the output.

Agent Modes

Amp has three agent modes:

  • Deep: Extended thinking via GPT-5.5 — maximum reasoning, higher latency and cost
  • Smart: SOTA models without constraints — best capability without extended thinking overhead
  • Rush: Fast mode, no reasoning — best for simple, quick tasks

Sourcegraph Foundation

Amp is built on Sourcegraph's code intelligence infrastructure. This means:

  • Strong codebase navigation and semantic understanding of large repos
  • Good at cross-repository tasks (following imports, understanding dependencies)
  • Integrates with Sourcegraph's existing enterprise code search

Model Support

Amp supports multiple backend models including GPT-5, Claude Sonnet 4, GPT Image 2, and others. Not locked to a single provider. This matters for teams that want to switch models for cost or capability reasons without changing tooling.

Pricing

Usage-based, with no markup on provider API costs for individuals. Minimum $5 credit to get started. Enterprise tier costs 50% more and adds SSO and zero data retention.

Strengths

  • Think-out-loud mode is genuinely useful for oversight
  • Sourcegraph code intelligence = strong large-codebase performance
  • Multi-model support
  • Good for code review and refactoring tasks where understanding the reasoning matters

Weaknesses

  • Smaller ecosystem and community than Cursor or Claude Code
  • Less mature hook/customization system than claude-code
  • Think-out-loud mode adds latency — not ideal for simple, fast tasks
  • Fewer integrations than VS Code / Cursor ecosystem

Use Cases

Best for: complex refactors where you want to supervise the reasoning, enterprise codebases with Sourcegraph already deployed, teams that want model flexibility.

Not ideal for: overnight-runs style unattended automation, quick completions, or workflows that need deep hook customization.

Related

claude-code · cursor · codex-cli · agentic-workflows

Sources