Amp

What is Amp?

Amp is a coding agent from Sourcegraph designed for repository work through terminal and editor interfaces. It can inspect code, use development tools, edit files, and maintain threads that developers can review or share. Amp should not be confused with Cody: official Sourcegraph documentation says Cody Free and Pro ended in July 2025, while Cody Enterprise remains available.

The product is useful to evaluate when a developer wants an agent rather than only completion. Its decision points are cloud processing, provider routing, command permissions, long-thread cost, and whether collaborative agent history improves review.

Terminal and thread workflow

Install an agent only after inspecting what it can access. Start Amp inside a defined repository and a clean branch. Give it acceptance criteria, relevant files, test commands, non-goals, and a clear boundary around credentials and network access. Ask for a plan before migrations or architectural changes.

Threads can preserve reasoning, tool activity, and output for collaboration. That history is valuable for review, but it may also contain source context, logs, prompts, or sensitive identifiers. Define who can access or share threads and how long they are retained. Remove secrets before they enter the interaction; later deletion is not a substitute for prevention.

Review each proposed command and resulting diff. Agents can install a mistaken package, alter a lockfile, run generated scripts, or expand scope after a test failure. Work inside a container or restricted development account where possible, and keep deployment and production infrastructure outside agent credentials.

Subscriptions and usage cost

The official Amp manual describes usage based on the actual models and tools involved, while the July 2026 subscriptions announcement adds monthly plans with included value. Pay-as-you-go remains an option. Because an agent can call several models or tools across a long thread, one user instruction is not one stable billable unit.

Track cost by accepted task. Separate exploration, code generation, repair loops, test execution, and discarded output. Compare a subscription's included usage with the same tasks on pay-as-you-go, then add developer review time. Set budgets before allowing large repository or background tasks.

Provider and subscription details are new and can change rapidly. Check current model routing, included usage, rollover or expiry, and overage behavior at purchase.

How to evaluate an Amp thread

Use the same bounded task with a fresh thread and with an established repository thread. The fresh run reveals how well Amp discovers context; the longer run reveals whether accumulated history improves consistency or adds irrelevant context and cost. Score clarification, plan quality, file selection, command safety, diff size, test usefulness, and whether the agent stops when evidence is missing.

Export or preserve the evidence needed for review without treating the agent's narrative as proof. A good summary should match the actual diff and command log. Re-run tests independently after the agent stops, because an agent can misread truncated output or report the last successful command while a later change invalidates it.

Privacy, providers, and security

Amp's security page says it does not clone or index the entire codebase as a permanent prerequisite, but task context still travels to the service and selected model providers. The page describes enterprise zero-data-retention arrangements with exceptions, along with the absence of self-hosting and some bring-your-own-key support.

If Amp connects to a personal ChatGPT, Grok, or other subscription, the third party's consumer terms can govern that processing. Do not assume Amp's enterprise promise extends to every externally connected model. Map each provider, retention period, training choice, region, and deletion process.

The terms prohibit sensitive data in the service and place responsibility for validating output on the user. Use sanitized evaluation repositories until the organization approves the full data path. Keep customer records, credentials, production logs, regulated data, and confidential documents out of prompts and tool output.

Teams should also define thread-sharing rules. A useful public thread can accidentally disclose repository paths, issue details, employee names, logs, or proprietary code. Review the complete thread—not only the final answer—before changing visibility or sending a link outside the organization.

Output review and alternatives

Amp can accelerate multi-file work, but it does not assume responsibility for correctness. Inspect code, license implications, dependencies, tests, security, and operational behavior. Run independent formatting, type checks, tests, SAST, secret scanning, dependency review, and manual authorization analysis before merge.

Choose Amp for a terminal-first agent with transparent usage concepts and collaborative threads. Compare Cursor or Windsurf when an AI-first editor is preferred, and GitHub Copilot when GitHub integration, completion, review, and organization policy should come from one platform.