Sourcegraph Amp

Sourcegraph Amp is the coding agent Sourcegraph built on top of fifteen years of code-search infrastructure. Unlike Copilot-style autocomplete or general-purpose agents, Amp is designed for engineers who work in million-file monorepos: it has live access to Sourcegraph's code graph — every reference, every call site, every dependency — runs in both VS Code and on the CLI, and uses frontier models with no per-message metering. Sourcegraph passes through token cost at-cost rather than capping prompts or charging a markup.

Amp plans multi-file changes, edits in place, runs commands, calls tools, and shares its thread state across desktop and CLI so you can hand off between environments. Start a refactor in VS Code, walk to a meeting, resume the same thread in the CLI on a server. The agent is grounded in the same code intelligence Cody uses, which is the practical advantage over Cursor and Copilot in monorepos: it actually knows the call graph rather than relying on embedding similarity.

The product has been adopted aggressively inside large engineering orgs that already pay for Sourcegraph Code Search and want one vendor for code intelligence, AI review (via Cody and Amp Review) and agentic coding. Sourcegraph also supports MCP in both directions: Amp can consume external MCP servers (filesystem, GitHub, browser tools), and Sourcegraph exposes the code graph itself as an MCP server so other agents like Claude Code, Cursor or OpenClaw can ground answers in private repos with permission awareness.

Pricing is consumption-based for individuals — you pay model token cost pass-through with no platform markup — and per-seat for enterprise teams. Sourcegraph's broader pricing surfaces Pro at $16/user/month plus Enterprise and Premium tiers for SOC 2, self-hosting, SSO and admin controls. The economics are unusual in a good way for heavy users: a developer who would burn through Cursor's $20 cap in a day pays only true model cost on Amp.

Real pros: monorepo grounding is genuinely better than any embedding-only competitor, the pass-through pricing rewards heavy use instead of capping it, and shared threads across VS Code and CLI map to how senior engineers actually work. Real cons: there is no managed model — you provide API keys and pay token cost, which is fine for power users but extra friction for managers buying for a team; the VS Code extension is newer than the Cody one; and the CLI is excellent but documentation lags the product.

If you are picking an agentic coding tool, compare Amp with Cody Sourcegraph (the chat/completion sibling), Cursor Agent for the most popular editor agent, GitHub Copilot Agents for the enterprise-default option, Windsurf for the alternative agent IDE, and Claude Code for terminal-first work that pairs nicely with Amp's MCP server. My recommendation: Amp is the right pick for senior engineers in large monorepos who want one platform for code intel and agentic coding without artificial caps.