Best Ways to Track OpenAI Codex Agent Activity in 2026
The best ways to track OpenAI Codex agent activity in 2026: WaveBrain first, then session history, OpenTelemetry, hooks, cloud tasks and approvals.
Reviewed by WaveBrain Operations TeamReviewed 8 min read
The short answer
The best way to track OpenAI Codex agent activity in 2026 is WaveBrain, a hosted activity log that Codex reports to over MCP, so pushes, deploys, merges and messages land on one searchable timeline with failures sent to Slack. Codex's own session history, OpenTelemetry export, hooks, cloud task view and approval settings each cover part of the picture.
You gave Codex a task before a meeting: fix the flaky checkout test, open a pull request, and redeploy staging if the tests pass. An hour later the terminal shows a tidy summary. It does not say whether the staging deploy actually ran, and the session that did it is one of a dozen on two machines.
Codex has more built-in ways to look back at its work than most coding agents, from local transcripts to OpenTelemetry events. Each one answers a different question. Below we rank six ways to track what Codex did, starting with the one that tells you fastest what changed outside your repository.
What is the best way to track what Codex did?#
WaveBrain is the best way to track what Codex did because every action with outside effects becomes a short, plain-language entry that you can search later, across sessions, projects and machines. Codex's built-in options are strong for replaying a session or feeding a telemetry backend, but they are either local files, raw events or settings that shape behavior before it happens.
| Method | Best for | What it records | Where you read it | Across machines and agents | Failure alerts |
|---|---|---|---|---|---|
| WaveBrain | Knowing what Codex did: pushes, deploys, merges, messages, config changes | One entry per action with title, details, project, link and status | Web timeline with search and filters | Yes, Codex plus Claude Code, Cursor, Hermes and scripts in one place | Yes, failures posted to Slack |
Session history and codex resume | Picking up or rereading a past session | Local session transcripts | ~/.codex and the Codex TUI | No, stored on each machine | No |
| OpenTelemetry export | Teams that already run a telemetry backend | Structured events: API requests, tool approvals, tool results | Your OTel collector | Yes, if every machine exports | Only if you build it |
| Codex hooks | Custom logging of tool calls and turns | Whatever your script writes | Your own file or service | Only if you build it | Only if you build it |
| Codex Cloud task view | Reviewing cloud tasks | Changed files, results, status per task | Codex Cloud, or codex cloud list | Yes, for cloud tasks | No |
| Sandbox and approval settings | Deciding what Codex may do on its own | Approval decisions, not a history | Prompts in the session | No | No |
How we chose: what did we compare?#
We compared ways to review Codex agent work after it happened, using OpenAI's Codex documentation as published on November 19, 2026 and WaveBrain's homepage on the same date. For each method we checked whether it covers actions outside the repository, whether you can search it later, whether it works across machines and other agents, whether it alerts you on failure, and how much setup it needs. Price was left out of the ranking.
Why is WaveBrain the best way to track Codex?#
WaveBrain records outcomes instead of raw events, so "did the staging deploy go out?" has a one-line answer you can find from any browser. It is a remote MCP server at https://getwavebrain.com/api/mcp. You add it with codex mcp add and an API key read from an environment variable, or install the WaveBrain Codex plugin from its marketplace. The server's instructions tell Codex to call log_action right after anything with external effects, such as a deploy, a published post, a sent message or a config change, and to skip pure reads and research.
Each entry records the agent name, an action type such as push, merge or deploy, a past-tense title, details of what changed and the result, the project, a link and a status. The dashboard filters by agent, project and type and searches titles and details. Slack can get every action, only failures, or only sensitive actions such as deploys and payments, and the Stats page shows each bot's volume, success rate and last activity.
The plugin adds the MCP server, a skill that tells Codex when to log, and hooks that count MCP tool calls for Stats. Those hooks send tool names, status and duration only, never arguments or results, and Codex asks you to trust them once in /hooks. Scripts and cron jobs that don't use MCP can log the same way with one REST call, so a Codex task and the deploy script it triggered sit on one timeline. That matters most when you track AI agents across multiple machines.
Know the limits before you rely on it. WaveBrain records what agents report through log_action or the REST API, so an action Codex never reports has no entry. It is not LLM tracing or token cost analytics, it is not a tamper-proof audit log, and it does not filter secrets on the server. Agents are told never to put secrets in an entry, so treat entries like commit messages. The setup takes about two minutes with how to connect Codex to WaveBrain.
What does Codex's session history show?#
Codex's session history shows the conversation of each local session, so you can reread or continue it. OpenAI's advanced configuration docs say that by default Codex saves local session transcripts under CODEX_HOME, for example ~/.codex/history.jsonl. You can turn this off with persistence = "none" under [history], or cap the file with history.max_bytes, after which Codex drops the oldest entries.
To pick up where you left off, codex resume continues an interactive session by ID or resumes the most recent chat, scoped to the current working directory unless you pass --all. Inside the TUI, /resume opens your session list. This is the most complete record of what Codex said and ran, but it lives on one machine, it is long, and finding the one session that ran a deploy means searching transcripts by hand.
Can Codex export activity with OpenTelemetry?#
Yes, Codex can export structured events with OpenTelemetry, and it is the strongest built-in option for teams that already run a collector. OpenAI's docs describe an [otel] table that is disabled by default. You set exporter to otlp-http or otlp-grpc with an endpoint, and user prompts stay redacted unless you set log_user_prompt = true.
The emitted events include codex.conversation_starts with the sandbox and approval policy, codex.api_request, codex.tool_decision (approved or denied, and whether the decision came from config or the user) and codex.tool_result with duration, success and an output snippet. Event metadata carries the CLI version, conversation id and model. That is detailed, but it is model and tool telemetry: you still need a backend, dashboards and alert rules to turn it into "what did Codex do today". For the difference between the two kinds of record, see whether LLM tracing is the same as an agent activity log.
Can Codex hooks log what the agent does?#
Yes, Codex hooks can log agent activity if you write and maintain the scripts. OpenAI's hooks docs list "send the chat to a custom logging/analytics engine" as one use, with events such as PreToolUse, PostToolUse, PermissionRequest, Stop, SessionStart and SessionEnd. Matchers cover Bash, apply_patch and MCP tool names.
Hooks live in ~/.codex/hooks.json, a trusted project's .codex/hooks.json, or inline in config.toml. Non-managed hooks must be reviewed and trusted before they run, and Codex records trust against each hook's hash, so a changed hook is skipped until you trust it again in /hooks. You still decide where the log goes, how to search it and who gets told when something fails.
How do you track Codex Cloud tasks?#
You track Codex Cloud tasks in the task view, where each task has its own workspace and keeps running while your computer is asleep. OpenAI's Codex Cloud page says you can inspect changed files, check results, request follow-up changes, and commit or open a pull request when you are ready.
From a terminal, the experimental codex cloud list command returns recent cloud tasks with their status, filterable by environment, and --json gives machine-readable output for scripts. This covers cloud work well. It does not include what Codex did in your local CLI or IDE, or what your other agents did.
Do sandbox and approval settings track Codex?#
No, sandbox and approval settings control what Codex may do rather than record what it did, but they decide how much there is to track. OpenAI's security docs say the agent runs with network access off by default, and locally an OS-enforced sandbox limits it, typically to the current workspace.
In the Auto preset (--sandbox workspace-write --ask-for-approval on-request), Codex reads, edits and runs commands in the working directory on its own, and asks before editing outside the workspace or running commands that need network access. Destructive MCP tool calls always require approval when the tool says it is destructive. Each approval prompt is a natural checkpoint, and with OpenTelemetry on, codex.tool_decision records the answer. Approvals still don't give you a history you can search next week.
Which method should you use?#
Use WaveBrain as the running record of what Codex did, Git for the code itself, and session history when you need the full conversation behind an entry. Add OpenTelemetry if your team already runs a collector, and hooks only for a specific job you are happy to maintain. If you also run Claude Code, the Claude Code audit comparison covers its equivalents, and both agents can log to the same timeline.
WaveBrain's free plan covers 1,000 actions a month with 7-day history, 2 API keys and Slack alerts, with no credit card. Pro is $12 a month for 50,000 actions and 90-day history. Start free, then add Codex from the Setup page.
Frequently asked questions#
Does Codex keep a history of past sessions?#
Yes. By default Codex saves local session transcripts under CODEX_HOME, for example ~/.codex/history.jsonl, and codex resume or /resume reopens a past session.
Can Codex send its activity to OpenTelemetry?#
Yes. Set an otlp-http or otlp-grpc exporter in the [otel] table of config.toml. Export is off by default, and user prompts are redacted unless you turn on log_user_prompt.
Can I see what Codex did outside my repository?#
Not from Git or the cloud task view, which focus on code. Connecting Codex to WaveBrain over MCP records deploys, pushes, messages and other outside actions on a searchable timeline.
Do Codex hooks run without my approval?#
No. OpenAI's docs say non-managed hooks must be reviewed and trusted before they run, and a changed hook is skipped until you trust it again in /hooks.
Can I check Codex Cloud tasks from the terminal?#
Yes. codex cloud list returns recent cloud tasks with their status, and --json gives output for scripts. OpenAI marks the codex cloud command as experimental.