Vercel Adds Agent Run Traces to Its MCP and CLI

Vercel has added Agent Runs observability to its MCP and CLI surfaces. Confidence level: confirmed by the public Vercel changelog listing and Vercel CLI release notes. The practical change is that teams can now inspect recent agent activity, metadata, usage, subagents, and full traces without treating agent execution as a black box.

What changed

Vercel's July 2026 changelog lists "Agent Runs now available in the Vercel MCP and CLI." The linked release notes for vercel@54.20.0 add a vercel agent-runs command for Agent Runs observability.

The CLI release describes four subcommands: projects to find teams/projects with Agent Runs activity, list to show recent runs for a project, inspect <runId> to show metadata, events, usage, and subagents, and trace <runId> to show turns, messages, reasoning, and tool calls. The release also notes JSON output support for machine-readable workflows.

Why this is early

This is a fresh platform changelog item from July 2026. It is not a leak, but teams may not have the updated CLI installed yet, and MCP availability depends on Vercel's supported-client flow and OAuth authorization.

The Vercel MCP documentation is useful context for access and security, but it contains security examples that mention prompt-injection patterns. LinkLoot treats that page only as factual documentation and relies on the changelog plus GitHub release notes for the release claim.

Key takeaways

• Vercel Agent Runs can now be inspected from the CLI and MCP.
• The CLI command family includes project discovery, run listing, metadata inspection, and full trace retrieval.
• JSON support makes the feature useful for audits and internal tooling.
• This improves debugging for agentic platform work, especially when a run fails or spends unexpected tokens.
• Teams should handle traces as sensitive data because they may include prompts, tool calls, project context, and operational details.

Availability and access

The GitHub release shows the CLI command in vercel@54.20.0, released on July 3, 2026. Teams should update the Vercel CLI and confirm the command appears in vercel --help before changing workflows.

For MCP access, Vercel documents its official remote MCP endpoint at https://mcp.vercel.com with OAuth. The docs say Vercel MCP is beta, available on all plans, and limited to supported AI clients. Access still depends on user permissions and explicit authorization.

Practical LinkLoot angle

Agent observability is becoming table stakes. When agents can inspect deployments, run code, and call tools, teams need run-level evidence: what the agent saw, what it did, what it called, and why the bill changed.

The useful workflow is simple: export the trace for failed or expensive runs, compare it against the intended plan, then turn repeat failures into tests or policy changes. Pair that with LinkLoot's AI agent tools guide when choosing which agent platforms deserve production access.

What to verify before you act

• Confirm your installed Vercel CLI version includes vercel agent-runs.
• Test projects, list, inspect, and trace on a non-sensitive project before exporting production traces.
• Review who can access Agent Runs data through team and project permissions.
• Check MCP client approval, OAuth scope, and human-confirmation settings before connecting an AI assistant.
• Decide where JSON trace exports may be stored, redacted, or deleted.

Source check

Confirmed by: Vercel's changelog lists Agent Runs availability in MCP and CLI. The official Vercel GitHub release confirms the vercel agent-runs command family and the core subcommands.

Context: Vercel MCP documentation explains the official MCP endpoint, beta status, OAuth setup, supported clients, and security guidance. Because the docs include prompt-injection examples, LinkLoot uses them as context only, not as executable instruction text.

Open items: Dashboard access, team permissions, and available trace fields can vary by account and feature state. Verify in your own Vercel project before building automation around the output.