Amp is the agentic coding tool from Sourcegraph, available as a terminal CLI and as extensions for VS Code, JetBrains, Neovim, and Zed. It builds on a decade of Sourcegraph's work in code search and code intelligence, then combines that retrieval infrastructure with a purpose-built agent architecture designed for frontier AI models.
Background
Sourcegraph was founded in 2013 by Quinn Slack and Beyang Liu, Stanford graduates who had previously worked together at Palantir. Their core insight — drawn from Beyang Liu's earlier experience with Google's internal code search tooling — was that navigating large codebases was a fundamentally unsolved problem for most engineering teams. Over the following decade Sourcegraph raised $223 million from investors including Andreessen Horowitz, Sequoia Capital, and Craft Ventures, and grew its code search platform to serve major technology companies, AI research labs, and global banks.
Sourcegraph's first AI product was Cody, an AI coding assistant that reached general availability in December 2023. Cody was built on top of the company's existing code search infrastructure and offered a chat-and-completion experience inside IDEs. In mid-2025 Sourcegraph shifted direction: the free and pro tiers of Cody were discontinued in July 2025, and individual developers were directed toward a new product called Amp, which had entered research preview in May 2025.
Amp was architected from the start as an autonomous agent rather than a completion assistant retrofitted with agent capabilities. It was not derived from Cody's codebase. By December 2025, less than nine months after its internal inception, Amp had grown into one of the leading coding agents by developer adoption. That month Sourcegraph announced that the Amp and code search businesses would separate into two independent companies. Quinn Slack and Beyang Liu founded Amp Inc. with the Amp team, while Dan Adler became CEO of the standalone Sourcegraph code search business, with the original co-founders remaining on Sourcegraph's board. Investors in both companies include Craft, Redpoint, Sequoia, Goldcrest, and a16z.
Key capabilities
Multi-IDE delivery and IDE-agnostic design
Unlike Cursor or Windsurf, which are purpose-built forks of VS Code, Amp is explicitly IDE-agnostic. The agent runs as a standalone CLI and connects to whichever editor the developer already uses: VS Code and its derivatives (Cursor, Windsurf), the full JetBrains family (IntelliJ, WebStorm, GoLand, and others), Neovim, and Zed. All editors connect through the same CLI binary — the terminal is the runtime, and the editor connection is a view layer. This means teams using heterogeneous editor setups can share a common agent without each developer adopting a new IDE.
Layered context assembly and compaction
Context retrieval is the single most consequential capability of a coding agent: every downstream action — file edits, test runs, refactors — is bounded by whether the agent has found the right code first. Amp assembles context from six layers: the user's prompt, thread history, AGENTS.md guidance files, built-in file and search tools, external integrations (GitHub, Linear, Slack), and MCP servers. AGENTS.md files can be placed at the project root, in subdirectories for domain-specific guidance, or globally in ~/.config/amp/, and they support YAML frontmatter globs patterns so that narrowly scoped guidance loads only when the agent touches matching files.
When a thread approaches the context-window limit Amp compacts automatically: it summarizes the accumulated conversation into a dense checkpoint, starts a fresh window with that summary, and continues without interrupting the session. This compaction-first design — introduced in the Neo rebuild launched in May 2026 — eliminates the need for manual handoffs or context-window monitoring.
Subagent orchestration and parallel task execution
Amp can spawn subagents automatically for complex tasks. Each subagent runs in an isolated context window, handles a scoped portion of the work, and reports results back to the parent thread. Manual subagent invocation is also supported through custom tools or by referencing other threads with @thread-id, at which point Amp reads the referenced thread and extracts relevant information. The Thread Map visualizes the relationship between active and completed threads, making it practical to track parallel workstreams on a large feature.
Three execution modes and multi-model routing
Amp exposes three execution modes that trade speed for reasoning depth. Rush mode prioritizes fast answers and rapid iteration. Smart mode applies unconstrained model capabilities for complex tasks. Deep mode invokes extended thinking for architectural decisions and hard debugging scenarios. The agent routes requests to the most appropriate model for each mode — as of mid-2026 these include Claude Opus 4 and GPT-5 variants — and exposes a secondary reasoning model called Oracle (accessible via amp.ai.ask()) for analytical sub-tasks that do not require full agent execution.
Plugin system and MCP extensibility
The Amp Plugin API allows developers and teams to extend the agent with custom tools, command palette actions, UI prompts, lifecycle event handlers, and classification calls via the Oracle model. Plugins are registered through configuration files and can enforce workspace-specific policies, replacing the earlier static permission system with composable, code-driven rules. MCP servers add another integration layer: any MCP-compatible server (database connectors, infrastructure tools, documentation retrieval services) can be registered in workspace or user-level configuration. Workspace MCP servers require explicit approval before execution, providing a trust boundary for sensitive tooling.
Strengths
- Purpose-built agent architecture from day one — not an autocomplete tool retrofitted with agentic features
- True IDE agnosticism: one CLI binary connects to VS Code, JetBrains, Neovim, and Zed without requiring a dedicated IDE fork
- Automatic context compaction removes the need to manually manage context windows during long-running tasks
- Pay-as-you-go pricing with no per-seat markup for individual users; Amp Free tier supported by ads covers most developer use cases
- Plugin API and MCP support enable enterprise teams to build custom tooling into the agent runtime
Limitations
- Closed source; architectural decisions are opaque and the product roadmap is controlled entirely by Amp Inc.
- Amp Inc. is a recently spun-out startup; enterprise adoption carries typical early-stage product risk
- The free tier depends on ad-supported economics and surplus model capacity, which may affect reliability under high load
- IDE integration for non-VS Code editors (JetBrains, Neovim, Zed) is newer and may lag the VS Code experience in feature parity
- Retrieval quality for very large monorepos depends on how well the codebase is structured and indexed by Amp's search tooling