Tool Introduction

Amp was created inside Sourcegraph, the code-intelligence company co-founded in 2013 by Quinn Slack and Beyang Liu, and launched as a research preview in May 2025. Sourcegraph lifted the waitlist on May 15, 2025, and introduced a no-cost tier called Amp Free on October 15, 2025.

On December 2, 2025, Amp spun out as an independent business, Amp Frontier Corporation, with Slack serving as chief executive officer and Liu among the founding roster. The company is headquartered at 455 Market St in San Francisco, California.

Demo Available?

No formal sales demo or time-limited trial is offered. Amp Free gives ongoing no-payment access to the full smart agent; paid usage starts at a $5 credit purchase.

Tool's Accuracy

Amp does not ship a proprietary model and does not publish a disclosed-methodology accuracy score such as SWE-bench Verified or Terminal-Bench for the product as it ships. Per the company's model-evaluation post, the team does not rely on formal quantified benchmarks to assess the frontier models it swaps into its deep, smart, and rush modes. Instead it uses qualitative, internal real-world usage testing and regression-style checks, stating that evals are primarily useful as unit tests and regression tests.

Aggregate Rating

4.5/5 stars (based on 90+ verified reviews across the internet)

Detailed Overview

What Sourcegraph Amp Does

Sourcegraph Amp reads a codebase, plans, edits files, and runs commands on the developer's behalf to finish coding work end to end. Work happens in threads that pair conversation with tool calls; when a thread reaches 90% of its context window, the agent compacts it automatically by summarizing and restarting. Developers install it through Homebrew or curl and drive it interactively or with an execute command.

Who Sourcegraph Amp Is For

The platform serves software developers who want an autonomous agent to handle multi-step work rather than line completions. Individual engineers run it on pay-as-you-go credits, while teams share a pooled credit balance across a workspace. Engineers tackling large-scale refactors turn to a context window of roughly 1M tokens, and organizations needing centralized cost tracking and access governance adopt its managed controls.

Top Differentiators

Sourcegraph Amp bills usage with no markup for individuals, passing provider API costs straight through: a thread using $2 of Anthropic and $0.50 of OpenAI API costs $2.50. Rather than committing to one model family, it routes between GPT-5.5 and Claude-family models to track whichever performs best. A curated toolset of Oracle, Librarian, Painter, and code review replaces an open marketplace of extensions.

Founding Year

2013

Target Industry

B2B, B2B2C, SMB, Enterprise, Software Development, Developer Tools

Available Integrations

VS Code: native integration once the Amp CLI is installed and the editor is running, giving Amp visibility into the open file and selection with full undo support; VS Code forks Cursor and Windsurf are also covered.
JetBrains: supported on IntelliJ, WebStorm, GoLand and others at versions 2025.1 and later, enabled by running amp --jetbrains.
Neovim: native terminal-editor support via a dedicated official Amp plugin distributed on GitHub, requiring a separate plugin install.
Zed: native command-palette integration after CLI install, following the same editor-running-then-run-amp pattern as VS Code.
GitHub: code-host connection via OAuth that powers the Librarian subagent, which searches public and granted private repositories on default branches only.
MCP servers: supports local command-based and remote HTTP-based Model Context Protocol servers, configured globally, per-workspace with approval, bundled in skills, or passed inline.
Linear: named MCP integration with automatic OAuth support, used for issue and ticket workflows.
Chrome DevTools: named MCP-based integration for browser and debugging workflows.
Custom plugins: first-party TypeScript extension mechanism that hooks lifecycle events and adds tools, commands, and subagents; global and workspace plugins are a limited experimental release.

Sourcegraph Amp uses a single pay-as-you-go credit system with no subscription tiers and no seat-based plans. Standard and team workspaces are charged zero markup on the providers' API pricing for actual LLM usage and certain tools like web search: a thread that incurs $2 in Anthropic API usage and $0.50 in OpenAI API usage deducts $2.50 from the Amp credits balance. The minimum credit purchase is $5, workspace credits pool among all members, and unused credits expire after one year of account inactivity. Invoices are processed through Stripe with VAT support.

Enterprise workspaces pay a 50% premium over the standard per-token rate; access requires a one-time $1,000 USD purchase that grants the workspace $1,000 USD of Enterprise usage credit plus enterprise features.

Amp Free is a no-payment tier: an hourly-replenishing free credit grant totaling roughly $10 per day, about $300 per month, giving access to the full smart agent powered by Opus 4.5 with GPT-5 and Gemini-3 subagents, plus a faster Rush mode driven by Haiku 4.5. Amp Free applies only to interactive and editor use; the CLI execute mode, other programmatic invocations, and the Amp SDK consume paid credits only. As of this check, ampcode.com/free states Amp Free is full for now, with new signups paused.

Pricing last verified: 07/04/2026
Disclaimer: Pricing is subject to change. Please confirm current pricing on the vendor site.

Features & Specs

Agent modes: deep (GPT-5.5 with extended reasoning for complex problems), smart (state-of-the-art models for maximum autonomy), rush (fast, low-token GPT-5.5 with no reasoning), plus an undiscoverable large mode for very large context.
GPT-5.5: 1M-token context window, priced at $5 per 1M input tokens and $30 per 1M output tokens, with reasoning levels low, medium, high, and xhigh.
Large Mode: powered by Claude Opus 4.8 as of a June 4, 2026 update, offering roughly 1M tokens of context for large-scale refactors.
Built-in tools: the Oracle second-opinion reasoning tool (Claude Fable 5 at high reasoning), the Librarian cross-repo search subagent, the Painter image tool (GPT Image 2, up to 3 reference images), and the amp review command.
CLI: interactive mode plus an execute mode (amp -x), streaming JSON output, stdin-driven multi-turn input, and inline MCP server declaration.
Editors: VS Code, JetBrains 2025.1 and later, Neovim, and Zed.
Pricing mechanics: $5 minimum credit purchase, workspace credits pooled across members, credits expire after one year of inactivity.
Security: TLS 1.2+ in transit, AES-256 at rest, SOC 2 Type II certification, and audit logging retained at least 30 days.

Security & Compliance

All traffic in transit is encrypted with TLS 1.2+, and all data stored by the Amp Server is encrypted at rest at the volume level using AES-256, including backups. Amp is SOC 2 Type II certified, with reports available through a Security Trust Portal, and runs annual penetration testing across clients, backend, and infrastructure.

Neither Amp nor its subprocessors train models on customer data unless explicitly opted in, and training cannot be enabled at all for Enterprise workspaces. Enterprise plans offer zero data retention, meaning LLM input and output is not retained by inference providers. For non-Enterprise use, data is deleted upon account deletion, and explicit thread deletion removes data within 30 days.

Authentication: Enterprise SSO and SAML via WorkOS with providers such as Okta, directory sync for automated provisioning and deprovisioning, and the option to require SSO as the sole login method.
Access controls: passkey authentication for sensitive and remote-control actions, thread visibility settings (private, workspace, group, unlisted), and IP allowlisting.
Additional controls: audit logging retained a minimum of 30 days, automatic secret redaction before data leaves the client, layered prompt-injection defenses, and a data-management API for enterprise security teams.

Tool's Alternatives

GitHub Copilot is natively built into GitHub, giving it reach into repos, pull requests, and Actions without leaving the platform teams already use for source control. It is most powerful when a team is already centered on GitHub as the code host, versus Amp's host-agnostic CLI-first design. It lets users assign tasks to agents like Copilot, Claude, and OpenAI Codex from one managed workspace.

Cursor offers an autonomy slider spanning Tab completions, targeted Cmd+K edits, and full autonomous Agent Mode, plus deep codebase-wide semantic indexing. It is fundamentally an IDE fork you adopt, a heavier commitment than Amp's bring-your-own-editor CLI model. Its agents can run on schedules and in parallel using Cursor's own compute, reaching into Slack and GitHub PR review.

Claude Code runs one underlying agent identically across terminal, VS Code, JetBrains, a desktop app, and the browser, with the same settings and MCP servers following the user everywhere. Most surfaces require a Claude subscription or Anthropic Console account. It adds built-in Routines, scheduled cloud-hosted recurring tasks, and native Slack, Telegram, and Discord channels for routing work into a session.

Cline is fully open source under Apache 2.0 with 250+ contributors and bring-your-own-key flexibility across 10+ providers including local Ollama and LM Studio. As a community project it lacks a centrally managed Enterprise tier with the governance controls Amp offers. It provides an explicit Plan-and-Act mode that shows the proposed strategy for review before execution.

Devin Desktop, formerly Windsurf by Cognition, positions itself as a multi-agent command center coordinating several coding agents side by side via the Agent Client Protocol. The standalone Windsurf brand has been absorbed and rebranded into Devin Desktop. It adds Spaces built on shared Git worktrees plus Kanban-style task boards for tracking concurrent agent runs.

Frequently Asked Questions

What is Sourcegraph Amp and who built it?

Amp is a multi-model AI coding agent created inside Sourcegraph and launched as a research preview in May 2025. It spun out as the independent company Amp Frontier Corporation on December 2, 2025.

How much does Sourcegraph Amp cost?

Amp uses pay-as-you-go credits with zero markup on provider API pricing. The minimum purchase is $5, and Enterprise access requires a one-time $1,000 USD purchase granting $1,000 of Enterprise usage credit at a rate 50% higher.

Which editors and platforms does Amp support?

Amp runs as a CLI installable via Homebrew or curl, with integrations for VS Code, JetBrains 2025.1 and later, Neovim, and Zed, plus remote agent control from a web or mobile session.

What are the deep, smart, and rush modes?

Deep uses GPT-5.5 with extended reasoning for complex problems, smart uses state-of-the-art models for maximum autonomy, and rush uses fast, low-token GPT-5.5 with no reasoning for small, well-defined tasks.

Does Amp offer a free tier?

Amp Free is a no-payment tier granting roughly $10 per day of hourly-replenishing credits with access to the full smart agent. As of this check, new signups are paused because the tier is full.

What security certifications does Amp hold?

Amp is SOC 2 Type II certified, encrypts traffic with TLS 1.2+ and data at rest with AES-256, runs annual penetration testing, and offers zero data retention on Enterprise plans.

Which models does Amp run?

Amp routes between GPT-5.5 with a 1M-token context window and Claude-family models, including Claude Opus 4.8 in Large Mode, Claude Fable 5 for the Oracle tool, and GPT Image 2 for the Painter.

What are the Oracle, Librarian, and Painter tools?

The Oracle is a second-opinion reasoning tool for debugging and review, the Librarian searches remote GitHub repositories on default branches, and the Painter generates and edits images from up to 3 reference images.

Comments are closed.

Sourcegraph Amp

Frontier coding agent built for leading models