An AI agent Vercel has built into its platform. Its defining trait is that it can use Vercel’s own context — deployments, logs, metrics, configuration, and connected repositories — directly as material, so it can investigate anomalies in production apps on its own and review pull requests. It was announced in October 2025, and chat and Slack integration have been added since. Any code fix it proposes is first run through the project’s real builds, tests, and linters inside a secure Sandbox, so off-target suggestions rarely slip through. As of August 2026 it is offered as a public beta for Pro and Enterprise teams.
Key Features
- Code Review (PR review): A review runs automatically when a pull request is created and when a batch of commits is pushed to an open PR. It reads the whole codebase as context rather than just the diff, and flags security problems, logic errors, and performance concerns
- Only Sandbox-validated fixes are shown: It generates a patch and then runs the project’s own builds, tests, and linters in a secure Sandbox. Only suggestions that pass those checks appear in the PR, where they can be applied with one click
- Investigations (automatic analysis of anomaly alerts): When an anomaly alert fires, it queries the logs and metrics around that moment, correlates patterns and related errors, and summarizes likely causes. Results appear under Observability → Alerts in the dashboard, and the analysis streams in real time while it runs
- Reads your existing AI rule files as-is: It automatically detects common coding-guideline files such as
AGENTS.md,CLAUDE.md,.cursor/rules/*.mdc, and.windsurfrules, and uses them as review context. Rules in parent directories are inherited, up to 50 KB in total - Callable from PR comments with
@vercel: Beyond automatic reviews, you can ask for work in a PR comment —@vercel run a review(re-review),@vercel fix the type errors(implement and commit a fix),@vercel why is this failing?(investigate). Replies land in the same thread - Read-only by default: When a task needs write access, it presents a scoped plan and waits for approval. After approval it either makes the change directly or opens a pull request, depending on the task
- Chat and Installation: You can ask questions in natural language from the dashboard or Slack, and hand off installing supported products such as Web Analytics and Speed Insights — from repository analysis through to opening a PR
Pricing
Vercel Agent is not charged per seat; it is usage-based, according to the tokens actually consumed. Provider inference is billed at cost with no markup, plus a “Vercel Token Rate” of $0.25 per million tokens.
| Feature | Included usage | Price beyond that |
|---|---|---|
| Chat (dashboard / Slack) | A limited number of simple requests during the beta | Inference at cost + $0.25 per million tokens |
| Code Review | None | Inference at cost + $0.25 per million tokens |
| Investigations | 10 per billing cycle with Observability Plus | Inference at cost + $0.25 per million tokens |
| Installation | Free | Usage charges for the installed products still apply |
Availability is limited to Pro and Enterprise teams (public beta). The Vercel Token Rate applies to input, output, and cached tokens alike, and is described as covering project context gathering, model routing, and the cost of running the infrastructure. Usage and cost can be reviewed per feature under Agent → Usage in the dashboard, where you can also configure auto-recharge and a monthly spending limit.
Pricing is current as of August 2026. Check the official site for the latest rates.
Pros & Cons
✅ Pros
- It can reach platform-side information — logs, metrics, deployment history, configuration — directly, so it starts with context that is hard to assemble with external tools
- Suggestions are narrowed to ones that passed real builds, tests, and linters, so there is far less of the noise typical of AI review
- No per-seat charge, so fixed costs do not jump as the team grows
- Existing
AGENTS.mdandCLAUDE.mdfiles work as-is, with no need to rewrite rules for the reviewer - Read-only by default with writes gated behind approval, which limits the risk of the agent changing things on its own
⚠️ Cons
- It assumes projects deployed on Vercel, so teams on other hosting get little benefit
- As of August 2026 it is a public beta and is not available on the Hobby plan
- Usage-based billing makes costs hard to predict when large PRs keep coming, so spending limits are effectively a prerequisite
- The free Investigations allowance requires an Observability Plus subscription and is modest at 10 per billing cycle
- Investigation output is a set of candidate causes; the final judgment and the fix remain the developer’s responsibility
Comparison with Similar Services
| Criteria | Vercel Agent | GitHub Copilot Code Review | CodeRabbit | Sentry |
|---|---|---|---|---|
| Main role | PR review + production investigation | PR review | PR review | Error monitoring and root-cause analysis |
| Runtime validation | Runs builds, tests, linters in a Sandbox | Mostly static analysis | Mostly static analysis | Centered on runtime error data |
| Access to production logs/metrics | Yes (Vercel Observability) | No | No | Yes (its own monitoring data) |
| Platform assumption | Deployed on Vercel | GitHub | GitHub / GitLab and others | Language- and platform-agnostic |
| Billing model | Token-based usage (no seat charge) | Copilot per-seat | Mainly per-seat | Usage based on event volume |
Who Is It For
- Teams already running production on Vercel that want incident investigation and code review handled within the same context
- Development teams that tried AI review before but abandoned it because there was too much noise
- Teams that want to avoid per-seat billing, especially where membership changes often
- Small operations teams that want a first read on the cause before they look at a late-night error spike alert
- Developers who already maintain coding conventions in
AGENTS.mdorCLAUDE.md
Summary
Vercel Agent connects two jobs that are usually handled separately — code review and production monitoring — through the context of a single platform. Narrowing suggestions through Sandbox validation is a practical answer to the reason AI review often gets ignored: too many comments to read. On the other hand, it is worth keeping in mind that it assumes Vercel, that it is still a public beta, and that it is usage-priced. If you are already on Vercel, an easy way in is to enable Code Review alone first, watch the cost and the quality of its comments, and expand to Investigations from there.