Sourcegraph Cody — The AI Assistant That Actually Understands Your Entire Codebase

Ask GitHub Copilot a question about your codebase and it gives you a decent answer based on the files you have open. Ask Cursor and it indexes your current project. Ask Sourcegraph Cody and it searches across every repository your organisation owns — hundreds of repos, millions of lines of code — and gives you an answer grounded in how your teams actually build software.

That difference matters more than most people realise.

The AI coding assistant conversation has been dominated by autocomplete speed and agentic flashiness — which tool writes code fastest, which one creates pull requests autonomously, which one has the best demo on Twitter. But for enterprise engineering teams working with large, sprawling codebases across dozens or hundreds of repositories, the bottleneck was never typing speed. It was understanding. Understanding where that authentication logic lives. Understanding why the payment service was built that way. Understanding how a change in one repository affects three others downstream.

Sourcegraph Cody is the AI assistant built specifically for that problem. And in 2026, with its pivot to enterprise-only, the launch of Deep Search, and the introduction of its sibling product Amp for agentic workflows, it deserves serious attention from any organisation where "how does our code actually work?" is a question developers struggle to answer.

What Cody Actually Does

Cody is an AI coding assistant built on top of Sourcegraph's code intelligence platform — the same platform that powers code search, navigation, and analysis for some of the largest engineering organisations in the world.

The core capability is context retrieval at scale. When you ask Cody a question or request a code change, it does not just look at your open file. It uses Sourcegraph's semantic code search to retrieve relevant code, patterns, and definitions from across your entire codebase — every repository, every branch, every language. It builds a context window from that retrieval and feeds it to the underlying LLM, producing responses that are grounded in how your organisation actually writes code.

In practical terms, Cody offers:

Inline code completion. Context-aware autocomplete suggestions as you type, informed by your project and (on Enterprise) your organisation's broader codebase. Available in VS Code, JetBrains IDEs, Visual Studio, Eclipse, Neovim, and Emacs.

AI chat. A conversational interface where you can ask questions about your code, generate functions, explain complex logic, write tests, and debug issues. The chat is connected to Sourcegraph's search, so answers reference actual code in your repositories rather than generic patterns from training data.

Deep Search. Introduced in early 2026, Deep Search uses a dedicated subagent that performs thorough, multi-step searches across your codebase to find files relevant to your query. It summarises findings for the main agent, saving context window tokens and enabling longer, more focused conversations. Deep Search also supports code navigation directly in the sidebar preview — interactive hover cards showing definitions, references, and type information without leaving the search context.

Context controls. You can @-mention specific repositories, files, symbols, and even web URLs to precisely define what context Cody should use. On Enterprise, you can search up to 10 repositories simultaneously. Context filters let administrators control which repositories and files Cody can and cannot access — critical for organisations with sensitive code that should not be exposed to AI.

The Context Engine — Why It Matters

Every AI coding assistant claims "codebase awareness." Here is how the approaches actually differ, and why Sourcegraph's matters for large organisations.

Cody employs a Retrieval-Augmented Generation (RAG) architecture that combines pre-indexed vector embeddings with advanced code-search capabilities. The system operates across three context layers:

1. Local file context — the immediate editor buffer and open files

2. Local repository context — the current codebase, indexed locally

3. Remote repository context — retrieved through Sourcegraph's Code Intelligence Platform, using search, code graph (SCIP), embeddings, and other relevance methods

This multi-layered approach means Cody understands symbol relationships, type hierarchies, call graphs, and dependency structures across your entire codebase. With context windows up to 1M tokens (using Claude Sonnet 4), it can ingest an enormous amount of relevant code.

Compare this to the competition: GitHub Copilot uses Copilot Spaces — curated collections that teams manually assemble. Effective, but the context is only as good as the curation effort. Cursor indexes your current project directory and uses `.cursorrules` files. Best in class for single-project context, but project-level only.

For a 5-person startup with one repo, this does not matter. For a 500-person engineering organisation with 200 repositories and services in four languages, it matters enormously. When a developer asks "how does the payment service handle refunds?", Cody searches the payment service, the shared libraries it depends on, the API gateway routing to it, and the integration tests verifying its behaviour — returning an answer grounded in all of that context. Copilot and Cursor cannot do this without significant manual setup.

Enterprise Features

Cody does not lock you into a single AI model. Enterprise customers can choose from Anthropic (Claude Sonnet, Opus), OpenAI (GPT-4o, o1), Google (Gemini 1.5 Pro, Flash), and Mistral — configured at the organisation level, switchable per task. Need a massive context window? Use Gemini's 1M+ tokens. Complex refactoring plan? Claude Opus. This flexibility future-proofs your investment as new models launch.

Beyond model choice, Cody Enterprise includes: self-hosted deployment (full control over data residency), SSO and SAML (all major identity providers), context filters (administrators define which repositories the AI can access), custom API keys (route inference through your existing provider agreements), audit and compliance (usage tracking for regulatory requirements), and integration with external tools like Notion, Linear, and Prometheus to bring documentation and observability data into the AI's context.

Amp — Sourcegraph's Agentic Layer

In 2026, Sourcegraph introduced Amp, an agentic AI tool designed for more complex, end-to-end development tasks. Where Cody excels at understanding and searching code, Amp is built for autonomously planning and executing changes.

Think of it this way: Cody helps you understand your codebase. Amp helps you change it. Cody tells you how the payment service handles refunds. Amp refactors the payment service to handle refunds differently.

Amp Enterprise requires a one-time $1,000 USD purchase that grants your team $1,000 USD of usage and upgrades your team to Enterprise. This positions it as a serious tool for teams already invested in the Sourcegraph platform rather than a casual experiment.

For organisations evaluating Sourcegraph, the combination of Cody (intelligence) and Amp (agency) creates a more complete AI coding platform than either product alone — though the total cost is higher than alternatives.

Cody vs Copilot vs Cursor — An Honest Comparison

Cody wins on codebase intelligence. If your engineering team works across many repositories, needs to understand legacy code, or spends significant time on "where does this logic live?" questions, Cody's integration with Sourcegraph's code search is unmatched. No other tool provides this depth of cross-repository understanding.

Copilot wins on breadth and ecosystem. It works in every major IDE, integrates with GitHub's entire platform (PRs, Issues, Actions), and its agentic capabilities are mature. For most teams, Copilot is the path of least resistance with excellent results.

Cursor wins on raw AI-assisted development power. Composer mode, parallel subagents, and deep project indexing make it the most capable tool for building and refactoring code. But it only works in its own IDE, and its codebase context is project-scoped, not organisation-scoped.

Pricing — What It Costs

There is no free tier and no individual plan. Sourcegraph discontinued Cody Free and Pro in July 2025. This was a deliberate strategic choice — Sourcegraph decided that Cody's real value lies in enterprise-scale codebase intelligence, not in competing with Copilot and Codeium on free-tier autocomplete.

They are probably right. Doubling down on "we understand your entire codebase better than anyone" is a strategy with genuine defensibility. But it means Cody is no longer a tool individuals can evaluate on their own. At $59/user/month — 3x Copilot Business — the value proposition must justify a steep premium. For organisations where codebase comprehension is a genuine bottleneck, it does.

Who It's For — and Who It's Not For

Use Cody if:

• Your engineering team works across many repositories and needs AI that understands the full picture — not just the current file or project
• You already use (or are evaluating) Sourcegraph for code search and navigation, and want AI capabilities integrated into that platform
• You are a large enterprise with a monorepo or complex microservices architecture where understanding code relationships across services is a daily challenge
• You need self-hosted AI coding assistance with control over which models you use and which code the AI can access
• Your developers spend more time reading and understanding code than writing it — Cody's retrieval-augmented answers save hours of manual code archaeology

Do not use Cody if:

• You are a solo developer, freelancer, or small startup — there is no free or affordable individual plan
• You primarily need fast autocomplete and do not care about cross-repository context — Copilot or Codeium are better value
• You want maximum agentic power — autonomous PR creation, multi-file refactoring, automated testing — Cursor and Copilot are ahead here
• Your team works on a single repository — Cody's multi-repo intelligence does not provide much advantage when there is only one repo to understand
• Budget is constrained — at $59/user/month, Cody costs roughly three times what Copilot Business charges

How to Get Started

1. Assess whether you have the right problem. Cody's value is clearest for organisations where codebase comprehension is a genuine productivity bottleneck. If your developers regularly struggle with "how does this system work?" or "where is this logic implemented?", Cody addresses that directly. If your team's bottleneck is writing speed, look elsewhere.

2. Start with Sourcegraph code search. If you are not already using Sourcegraph, evaluate the code search platform first. Cody's intelligence is built on top of Sourcegraph's indexing — the AI is only as good as the code graph it draws from.

3. Engage Sourcegraph's enterprise team. There is no self-serve sign-up for Cody Enterprise. Contact Sourcegraph to scope your deployment, discuss model preferences, and configure context filters for sensitive repositories.

4. Pilot with one team. Do not roll out to the entire engineering organisation at once. Choose a team that works across multiple repositories, give them Cody for a month, and measure impact on time-to-resolution for code comprehension tasks, onboarding speed for new team members, and code review quality.

5. Compare against Copilot Business. Run Cody alongside Copilot for your pilot team. If the multi-repository context does not provide meaningful value above what Copilot offers, the 3x price premium is not justified. Be honest about the results.

The Bigger Picture

The AI coding assistant market is splitting into two tiers. The first tier — Copilot, Cursor, Windsurf, Codeium — competes on autocomplete quality, agentic capabilities, and price. These tools make individual developers faster at writing code.

The second tier — where Cody sits — competes on codebase intelligence. These tools make entire engineering organisations faster at understanding, navigating, and maintaining complex systems. Writing new code is only part of what engineers do. Reading existing code, tracing dependencies, understanding why something was built a certain way, figuring out the downstream impact of a change — that is where most engineering time actually goes.

Cody is the best tool in the market for that second problem. It is not the cheapest. It is not the flashiest. It does not have the most impressive demo. But for enterprises where "understanding the codebase" is measured in engineer-hours per week, Sourcegraph Cody turns a manual archaeology expedition into a conversation.

That is worth paying for — if you have the right problem.

Digital by Default helps enterprises evaluate AI development tools for complex engineering environments. If you are weighing Cody against Copilot or Cursor for a large codebase and want an honest assessment, [get in touch](/contact).