JetBrains has released an experimental AI Agents plugin for TeamCity that runs OpenAI’s Codex and Anthropic’s Claude Code inside CI builds, rather than only on a developer’s laptop. In a TeamCity blog post dated 16 November 2026, the company says the plugin can install the chosen agent on the build machine, hand over provider credentials only for the duration of the build, and record token use and the agent’s full trajectory. The plugin is available from JetBrains Marketplace. JetBrains calls it experimental, and says its features may change.
Quick Take
- What launched: An experimental TeamCity plugin that runs Codex or Claude Code as part of a CI/CD build.
- What it automates: agent install, a project-level connection to the model provider, short-lived credentials, token statistics, and a trajectory log in HTML and JSON.
- Supported agents: Codex and Claude Code. Supported connections: JetBrains Air, OpenAI, Anthropic, LiteLLM and OpenRouter.
- How you call it: from an ordinary TeamCity Command Line build step, for example “codex exec –ephemeral –skip-git-repo-check” or “claude -p”, with your own prompt.
- The caution: trajectory artifacts may contain prompts, source code and model output, and should be treated as sensitive. The plugin is not presented as ready for a release-critical pipeline.
In plain English
JetBrains has made it easier to let a coding agent work inside your build server, not just on someone’s computer. The new TeamCity plugin can install Codex or Claude Code, give the build a temporary key, run a prompt you write, and then show you what the agent did and how many tokens it used. It is officially experimental. That makes it useful for trying things such as summarizing a repo or looking at a failed build, and a poor fit for any pipeline your release depends on, at least for now.
What happened
The official TeamCity blog, in a post dated 16 November 2026 and written by Olga Bedrina, says “AI coding agents don’t have to run only on a developer’s machine.” The company lists review, investigating failed builds, updating documentation and “other repository tasks” as examples of work that can move into CI.
The plugin’s job is the dull part of that setup. When a build starts, it checks whether the selected client is on the build agent and installs it if needed. Credentials live as a TeamCity connection at project level, not as a permanent secret on each machine. “The plugin provides the necessary credentials to the TeamCity agent only for the duration of the build and removes them afterward.” After the run, it publishes prompt, completion, cached and total token counts as ordinary TeamCity statistics, so teams can chart usage and spot spikes.
It also records the agent’s trajectory. HTML gives “a readable view of the agent’s messages, tool calls, and token usage.” JSON is meant for looking at many runs at once, “to understand why an agent behaves a certain way or how to adjust the original prompt.” The files are stored as hidden build artifacts under “.teamcity/ai-agents/”. JetBrains shows how to surface the HTML report as a “AI Agent Trajectory” tab on the build results page.
How you set it up, and what to watch
Setup has three steps in the official post: create a connection under Project Settings, add the AI Agent build feature and choose Codex or Claude Code, then add a Command Line step with your prompt. The examples are a repository summary, not a production deploy. Gateways such as OpenRouter and LiteLLM are supported, so the model behind the agent does not have to come directly from OpenAI or Anthropic. JetBrains Air can use a service-account client ID and secret, “which makes it possible to track shared AI usage and spending.”
Two warnings in the post matter more than the feature list. First, “the generated artifacts may contain prompts, source code, and model output. Treat them the same way you would other sensitive build artifacts and make sure access is appropriately restricted.” A CI log that includes the agent’s chain of thought can leak secrets just as easily as a chat transcript. Second, “the plugin is currently experimental, and its functionality may change as we continue developing it.”
That combination, a real CI hook plus an explicit experimental label, is the right way to ship this kind of tool. It also means teams should not point the agent at deploy keys, production credentials or unreviewed code changes without a human gate.
Why it matters
Coding agents are moving off the laptop and into shared systems. That is already visible in product features such as Claude Code’s experimental Code Mode and in teams that already run agents hard, including the Grok Bot team’s month of 11,000 pull requests. Putting Codex or Claude Code in TeamCity is a smaller, more concrete step: the agent becomes another build step, with token bills you can chart and a trajectory you can inspect.
The security model is the part to take seriously. Short-lived credentials are better than leaving an API key on a reused build agent. Trajectory files, though, create a new pile of sensitive artifacts. If those files are too widely visible, the plugin’s observability becomes a leak. That is not an argument against trying it. It is an argument for treating the output like source, not like a log anyone can browse.
Cost control will follow. Once agents run unattended in CI, token use can jump without a developer noticing. The plugin’s statistics help. So does a clear-eyed view of cost claims in this market, as we found when Reason Machines said its coding harness could cut costs by 20 times.
What’s still open
- Which TeamCity editions and versions are supported, and whether the plugin will stay experimental or move into a supported product.
- How the install step behaves on locked-down or air-gapped build agents, and what it downloads to run Codex or Claude Code.
- Whether JetBrains will add other agents beyond the two it supports today.
- How teams should retain or purge trajectory artifacts under their source-code and compliance rules.
Sources
- JetBrains TeamCity blog, Olga Bedrina, “Run AI Agents in TeamCity,” 16 November 2026
i10x Perspective
The useful idea in this plugin is not “AI in CI,” which was already possible with a shell script and a key. It is the boring machinery around it: temporary credentials, token charts and a trajectory you can read after the build goes wrong. That is how agent features become operations. Keep it in a trial project until JetBrains takes the experimental sticker off, and assume the HTML report is as sensitive as the repo it came from.






