Sourcegraph's AI coding assistant, now sold only on the Enterprise plan: whole-repo context from Sourcegraph's code search, on an annual per-seat contract.
Last updated
Before anything else about the product: Cody is an enterprise purchase now, and only an enterprise purchase. Sourcegraph closed new signups for Cody Free, Cody Pro and Enterprise Starter on 25 June 2025 and shut those plans down on 23 July 2025, announced on its own blog under the title "Changes to Cody Free, Pro, and Enterprise Starter plans". The consumer-facing marketing page at sourcegraph.com/cody now redirects to the documentation, and the first line of that documentation is a support notice: Sourcegraph Enterprise. The listed price is 59 dollars per user per month on an annual contract.
So if you arrived here as an individual developer comparing assistants, this one is not on the shortlist any more, and no amount of feature discussion changes that. Everything below is written for the only reader who can still act on it: someone deciding whether an organisation should run Sourcegraph and put Cody on top of it.
Sourcegraph sells code search to organisations whose codebases are too large for anyone to hold in their head — hundreds of repositories, millions of files, years of accumulated decisions. Cody is what happens when you put a language model in front of that index. Every assistant has to answer the same question, which is which code the model should actually see, and most answer it by guessing from your open tabs and recent edits. Cody answers it by running a search.
That is an architectural difference rather than a marketing one, and it has a consequence worth understanding before a trial: Cody's value is not roughly constant across teams the way GitHub Copilot's or Cursor's is. It scales with the size and the disorder of the thing being searched. It runs as an extension in VS Code and the JetBrains IDEs and does the usual things — chat, inline completion, commands for drafting tests or explaining a selection — and nobody picks it for those. What differs is that it can pull in code from repositories nobody on the team has open, because Sourcegraph already indexed them. Ask why a service returns a particular error and a local-context tool can only reason about the file in front of it; a search-backed one can surface the other call sites, the shared helper living two repositories over, and the migration that changed the behaviour in the first place.
Cody also lets an organisation choose among several underlying models rather than being welded to one vendor's. That reads like a checkbox and is not one: where legal has approved exactly one model provider, an assistant that cannot switch providers is an assistant that cannot be deployed at all.
59 dollars per user per month, committed annually, is roughly three times what the per-seat assistants cost, and that is only the licence. The context that differentiates Cody depends on Sourcegraph's index, so the actual decision is whether the organisation is prepared to run Sourcegraph — self-hosted, dedicated, or managed — and to keep it indexed as the code moves. For an enterprise already running it, Cody is close to free marginal effort, and the security posture is a genuine argument in its favour: code is indexed and served inside infrastructure you control rather than shipped to a shared third-party service, which is the kind of guarantee that decides procurement in regulated environments. Teams working through what those guarantees are actually worth may find the confidential-computation framing a useful companion. For a team not already running Sourcegraph, the honest cost of Cody includes standing up and operating a code search platform, which is a different conversation from installing an extension.
Retrieval only helps when retrieval is hard. On a project of a few dozen files, a modern assistant can hold the relevant code in its context window directly, and the machinery that makes Cody interesting has nothing left to do. You still pay for it, though — an index to stand up, a deployment to configure, and freshness to worry about, because a search index that lags behind the branch you are working on will describe code that no longer exists with complete confidence.
The inverse case is where the price makes sense. A codebase with undocumented internal libraries, two half-finished framework migrations, and conventions that live only in the memory of people who left is a retrieval problem, and retrieval is what Sourcegraph spent years building. So the question to settle first is not whether Cody is good. It is whether your estate is large enough and tangled enough that whole-estate context justifies both the seat price and the platform beneath it.
If you are one developer, or a startup, or anyone without a procurement process, pick something else — there is no tier for you, and that is the end of the analysis. If the goal is to hand over a ticket and return to a finished pull request, Devin is aimed at that job in a way Cody is not; Cody is a strongly grounded assistant rather than an autonomous one. If you want the most fluid in-editor experience with no infrastructure behind it, the Cursor and Copilot comparison is the more relevant read. If the hard requirement is that no source code may leave your network under any circumstances, an on-premises completion tool like Tabnine attacks that directly rather than as a side effect of where your search index happens to live. And if you are starting a greenfield project this week, there is nothing to search yet — revisit it when the codebase has enough history to be worth indexing.
A new engineer can ask in plain language how a subsystem works and get an answer grounded in the code that is actually deployed rather than documentation written two reorganisations ago. This is the single most commonly cited reason large engineering organisations keep paying for it, and it is the use case that most obviously justifies an enterprise seat price.
Platform and infrastructure teams use cross-repository context to find out who consumes an interface before they change it. An assistant limited to the open workspace structurally cannot answer that question, because the consumers are in repositories nobody has open.
In an old codebase the right answer is usually not the idiomatic answer from a training set — it is whatever the team already does. Search-backed context lets the assistant locate an existing implementation of the same pattern, so new code matches house convention rather than a generic blog post.
Where code cannot be sent to a shared cloud service, running Cody against a self-hosted or dedicated Sourcegraph deployment keeps indexing and retrieval inside a controlled boundary, and the ability to select among approved model providers matters as much as the deployment topology to the review board signing it off.
There is no individual tier. Cody Free, Cody Pro and Enterprise Starter stopped accepting new signups on 25 June 2025 and were discontinued on 23 July 2025; what remains is Sourcegraph Enterprise, listed at 59 dollars per user per month on an annual commitment. Any page still describing a free or 9-dollar Pro plan — and there are many — predates that change. Budget for the platform as well as the seats: the recurring cost of Cody is not only its licence but the Sourcegraph deployment that gives it something to search, plus whatever the chosen model provider charges under the terms you negotiate. Sourcegraph has since put its newer effort behind a separate product called Amp, which is sold on its own terms rather than as a Cody tier, so confirm which product a quote actually covers before signing.
No. Sourcegraph closed new signups on 25 June 2025 and ended Cody Free, Cody Pro and Enterprise Starter on 23 July 2025. Cody is supported on Sourcegraph Enterprise only, and the documentation says so in its first line. If a comparison table still shows a free Cody tier, it has not been updated since mid-2025.
Something with a tier they can actually buy. Copilot and Cursor both sell to individuals; Tabnine is the answer when code must not leave your machine; Ollama plus an editor plugin is the answer when you want to pay nothing and own the hardware. None of them reproduce Cody's cross-repository retrieval, but that capability was never really aimed at a single developer's project anyway.
Where the context comes from. Copilot and Cursor build context primarily from your open workspace and are generally faster and smoother for moment-to-moment editing. Cody queries Sourcegraph's index, which lets it answer questions that span repositories you do not have open. On a small project that difference is invisible; on a large multi-repository estate it is the whole point, which is also why the product ended up priced and packaged for that estate and nobody else.
Yes, and now literally so — Cody is only supported on Sourcegraph Enterprise. The whole-codebase and cross-repository context that distinguishes it comes from Sourcegraph's search index, so the adoption decision is really about whether you are willing to run and maintain that platform. Teams that want context with no infrastructure step should look at editor-native tools instead.
Yes — running it against a self-hosted or dedicated Sourcegraph deployment is one of the main reasons security-conscious organisations still choose it over cloud-only competitors. Confirm the current deployment options and the data-handling terms for whichever model provider you select, since those are two separate questions and only one of them is answered by self-hosting the index.
Model selection is part of the enterprise offering, which matters more for compliance than for quality. Organisations frequently have an approved list of model providers, and being able to point the assistant at an approved one is often the difference between a rollout and a rejected proposal.
Not in the sense that Devin is. It is best understood as a well-grounded assistant that answers questions and drafts changes with unusually good knowledge of your system, rather than something you assign a ticket to and leave alone. If autonomous execution is the requirement, evaluate agent products against that requirement directly instead of expecting Cody to grow into one.
Full review coming soon.