Field noteNov 16, 20265 min read1 source

JetBrains releases experimental TeamCity AI Agents plugin to run Codex and Claude Code in CI builds

In one minute

JetBrains has released an experimental TeamCity AI Agents plugin that runs Codex and Claude Code within CI builds. It handles temporary provider credentials, usage statistics, and agent trajectories, making it possible for teams to test coding-agent workflows in CI with operational visibility built in.

What changed: JetBrains introduces an experimental TeamCity plugin to run Codex and Claude Code in CI builds, with provider connections, credential handling, and usage tracking.

What happened

JetBrains says its experimental AI Agents plugin lets teams run Codex and Claude Code directly in TeamCity builds. The stated use cases include reviewing code, investigating failed builds, updating documentation, and other repository tasks. The plugin installs the selected coding-agent client on a TeamCity build agent if it is not already present. A team can then invoke codex or claude from a regular Command Line build step.

The plugin connects to an LLM provider through a TeamCity connection. JetBrains lists OpenAI, Anthropic, OpenRouter, LiteLLM, and JetBrains Air as supported providers. The company notes that gateways such as OpenRouter and LiteLLM can expose models beyond those supplied directly by the named services. Teams configure the connection in Project Settings, add the AI Agent build feature to a build configuration, choose Codex or Claude Code and the provider connection, and add a command-line prompt. JetBrains’ examples are codex exec --ephemeral --skip-git-repo-check "summarize this repository" and claude -p "summarize this repository".

Credential handling is one of the plugin’s explicit functions. JetBrains says provider credentials are managed through TeamCity connections rather than permanently stored on individual build agents. The plugin provides credentials to the agent only for the duration of a build and removes them afterward. This addresses a risk JetBrains identifies in reused build agents, where credentials left on a machine could persist across builds.

The plugin also publishes agent-run statistics as TeamCity statistical values: prompt, completion, cached, and total token counts. Teams can use these values to chart AI usage over time and investigate unexpected consumption changes. It records agent trajectories in HTML and JSON. The HTML report presents messages, tool calls, and token usage for inspection; JSON supports programmatic analysis across runs. Trajectory files and telemetry are published as hidden build artifacts under .teamcity/ai-agents/. A TeamCity report tab can expose the HTML trajectory report in build results.

JetBrains describes the plugin as experimental and says its functionality may change. The announcement does not state a price or provide performance or cost benchmarks. It warns that generated artifacts may contain prompts, source code, and model output, and advises treating them as sensitive build artifacts with appropriately restricted access.

What the archive adds

New Runtime’s earlier coverage on 2026-02-05 described GitHub’s Agent HQ preview, where users could select Codex, Claude, and Copilot within GitHub and VS Code. This is a new event: JetBrains brings two coding agents into TeamCity’s CI build workflow, with credential scoping and trajectory capture, rather than adding them to an interactive coding environment.

Trend signal

This supports agent security moves to runtime boundaries: JetBrains’ design treats credential lifetime and the build environment as controls around agent execution, rather than relying only on instructions in a prompt. The plugin’s trajectory artifacts also make actions inspectable after a run, although the announcement does not establish how comprehensive those controls are in practice.

Why now

JetBrains frames the launch around coding agents becoming useful beyond a developer’s local machine. Its plugin makes them callable from existing CI build steps, where teams already run repository tasks such as tests and build checks. The experimental label signals that this is an early integration rather than a mature, generally available TeamCity feature.

Why it matters

CI can give teams a repeatable place to trial agent tasks against repositories, but repeatability is not enough if credentials persist on shared runners or the agent’s actions cannot be examined. JetBrains’ approach combines three useful operating controls: build-scoped provider credentials, token statistics attached to TeamCity, and saved trajectories that show messages and tool calls. Together, they make it easier to investigate what an agent did and compare usage across runs.

These controls do not make every CI task safe to automate. Teams still need to decide which repositories and build configurations may run agents, which outputs require human review, and who can read trajectory artifacts. In particular, the artifacts may contain source code and prompts. A cautious pilot can therefore use a narrow task and tightly restricted artifacts before expanding agent access across builds.

Limits and uncertainty

The plugin is experimental, and JetBrains says its functionality may change. The announcement does not quantify reliability, latency, or costs, nor does it report results from deployments. It does not describe a permission model for what Codex or Claude Code can do within a build, or specify safeguards for agent-authored changes. Supported providers are named, but the announcement does not detail every model or configuration available through them. Teams should validate credential cleanup and artifact access in their own TeamCity setup rather than assume the announcement establishes security outcomes.

Skill to build now

Set up a narrowly scoped TeamCity build configuration for one low-risk repository task. Store the provider credential as a TeamCity connection, call the agent in a Command Line step, and publish its trajectory report only to a restricted audience. Compare token statistics across several runs and review the trajectory before deciding whether the task is suitable for a broader rollout.

Reader takeaway

The useful change is not simply that coding agents can run in CI: TeamCity also captures their trajectories and token usage while limiting provider credentials to the build. Treat it as an experimental pilot, keep the task narrow, and restrict access to artifacts that may expose prompts or source code.

Evidence / sources

  1. 1blog.jetbrains.com/teamcity/2026/11/run-ai-agents-in-teamcity
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02topicCodexExplore the codex topic hub.
  3. 03topicClaude CodeExplore the claude-code topic hub.
  4. 04archiveField NotesOpen the latest editorial analysis.
  5. 05sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract