· Clemens
Give a pull request to someone who joined last week and you get style notes. Give it to the person who has been on the team five years and you get "careful, checkout reads that field". Same diff, different amount of stuff already in their head.
Most AI code reviewers are the person who joined last week. They see a patch, a bit of the file around it, and nothing else. So you get a comment about a missing docstring while the thing that will page you at 3am sits two files over, in a caller nobody opened.
We built Context Goblin to be the other one.
Most of the work happens before the model sees anything #
When a PR opens, the first part of the run goes into working out what the change actually touches. The agents only start reading once that's done.
The context is assembled from a few places.
The pull request itself, which is the part everyone does: diff, commits, the review conversation so far.
The code around it. We build a graph of your repo with codebase-memory-mcp and use it to find the code that reaches the code you changed. Real callers, not name matches.
We keep a map of how your repositories are wired together: the calls between services and the dependencies between them. The part of that map sitting next to your diff comes along into the review.
You can also plug in your own MCP servers, so the agents read your tracker or your wiki alongside the code. That shifts the question from "is this code correct" to "does this do what the ticket asked for". Credentials stay in Google Secret Manager and are resolved per organisation for that one run.
The context then goes to several sub-agents, each looking at the change from a different angle. A coordinator merges what they found and posts a single review. Your code is cloned for that one run and deleted when the review is posted, and none of it is kept or trained on.
The honest part #
Plenty of tools in this space say they assist your reviewers. We're not going for that. What we want is for nobody to open the PR a second time. The review lands, it's right, the change ships.
That's on us to earn, not on you to work around. A reviewer you have to double-check costs you the same attention as doing the review yourself, plus the wait. So findings get checked against the code before they're posted, the bot says nothing when it has nothing, and it won't speculate about a caller it couldn't resolve in the graph.
We're not there yet. The cross-repo map is only as current as the last time we looked at your repositories, and some of what it knows is inferred rather than proven. Making every one of those findings traceable back to something you can check yourself is what we're building next.
We're also not showing you a benchmark yet. We'd rather run on other people's repositories first and then publish what actually came out, findings and misses both, once there's enough of it to be worth reading.
Trying it #
Install the GitHub App, pick your repos, and reviews show up on new PRs by themselves. @-mention the bot if you want one re-run. The same engine ships as an MCP server if you'd rather drive it from your own agents.
Access is manual at the moment. You fill in a form, we look at it, we let you in. Usually within a day. We know that costs us signups. We'd still rather bring people on slowly and watch how the first reviews land on their repos.
Context Goblin came out of the 2026 Bitmovin AI Incubator, which is a nicer way of saying it started as an internal thing for reviewing our own PRs and got good enough that we pointed it outward.
If you want in: contextgoblin.com. And if you think the whole premise is wrong, say so. We'd like to know which reviewer you're comparing it against.