Debug Vercel agent runs from MCP or the CLI
Vercel now exposes eve Agent Runs through MCP tools and CLI commands, so teams can inspect traces, reasoning, tool calls, token usage, lifecycle events, and subagent data without staying inside the dashboard.
Vercel has added Agent Runs access through both the Vercel MCP and Vercel CLI for eve, its open-source agent framework. Confidence level: confirmed. The release gives agents and developers a direct way to inspect traces, reasoning, tool calls, token usage, lifecycle events, and subagent data without treating the dashboard as the only debugging surface.
What changed
On July 3, 2026, Vercel announced four MCP tools and matching CLI commands for Agent Runs. Vercel says eve traces are automatically ingested when deployed to Vercel and become available as Agent Runs.
The new surface covers project discovery, recent run lists, run metadata, lifecycle events, usage, subagent data, and full trace retrieval. The CLI path matters because every subcommand supports --json, and traces render as Markdown when piped.
| Need | MCP tool | CLI command | Useful output |
|---|---|---|---|
| Find projects with agent activity | list_agent_run_projects | vercel agent-runs projects | Teams and projects with run data |
| List recent runs | list_agent_runs | vercel agent-runs list | Recent Agent Runs for a project |
| Inspect one run | get_agent_run | vercel agent-runs inspect <runId> | Metadata, lifecycle, usage, subagents |
| Retrieve trace | get_agent_run_trace | vercel agent-runs trace <runId> | Turns, messages, reasoning, tool calls, token usage |
Key takeaways
- Agent Runs are now reachable from Vercel MCP tools and the Vercel CLI.
- The feature is tied to eve traces deployed on Vercel.
- CLI output can be machine-readable with
--json. - Piped traces render as Markdown, which makes review artifacts easier to save and share.
- The practical value is debugging and cost visibility, not just another dashboard view.
Availability and access
Vercel says users can get started by installing the Vercel MCP with npx add-mcp https://mcp.vercel.com or by upgrading the CLI with npm i -g vercel@latest.
The announcement does not list a separate price for Agent Runs access. Teams should still check their Vercel plan, eve deployment setup, and any internal policy around exposing agent traces, because traces can include reasoning, tool input and output, and token usage.
Practical LinkLoot angle
Agent observability is becoming part of the build loop. If an agent changes production code, retries a failing task, burns tokens, or calls the wrong tool, a team needs a trace it can inspect outside a pretty dashboard. Vercel's CLI path makes that trace scriptable.
For a practical workflow, ask the agent to fetch its latest run, export the trace to Markdown, and summarize failure points with exact tool calls and token spikes. That fits the same pattern LinkLoot tracks in /guides/ai-agent-tools: useful agent tooling should make behavior inspectable, not just autonomous.
What to verify before you act
- Confirm the project uses eve and deploys traces to Vercel.
- Upgrade the Vercel CLI before relying on the new
agent-runscommands. - Review whether traces include sensitive prompts, tool inputs, outputs, or customer data.
- Test
--jsonoutput before wiring it into CI, cost alerts, or agent self-debugging. - Decide retention and sharing rules for Markdown trace exports.
Source check
Confirmed by: Vercel's changelog confirms the MCP tools, CLI commands, trace fields, --json support, Markdown rendering when piped, and install/upgrade commands.
Context: Vercel's changelog index corroborates the July 3 entry. Independent developer commentary treats the release as an agent-observability improvement, but the factual claim rests on Vercel's own announcement.
This post does not claim broad market adoption, pricing changes, or non-eve support beyond what Vercel published.
Vercel added MCP tools and CLI commands to find projects, list runs, inspect one run, and retrieve full traces for eve Agent Runs.