Vercel Eve, GitHub Agent Finder, and Strapi's MCP Server — AI Digest for June 18, 2026
Six builder-relevant updates today: Vercel open-sourced eve for filesystem-first agents; GitHub shipped Agent Finder and repo-reading CLI commands; JetBrains moved Junie out of beta; Strapi added a built-in MCP server; 42Crunch brought API-security checks into Copilot; and Google's consumer Gemini CLI path now needs migration checks.
The useful pattern in today's AI tooling news is not a bigger model. It is packaging: agents are getting registries, file layouts, permission scopes, terminal primitives, and security checks that make them easier to wire into real projects.
| What changed | Current status | Why it matters for builders |
|---|---|---|
| Vercel introduced eve, Vercel Agent, Vercel Connect, and full-stack service support at Ship. Vercel says agent-triggered deployments rose from less than 3% to more than half of deployments over six months, while AI Gateway token volume rose from roughly 2 trillion to 20 trillion tokens per month. Vercel announcement | eve is open source and in beta; the GitHub repo describes it as a filesystem-first framework for durable AI agents and shows Apache-2.0 licensing. Vercel eve GitHub repo | If you already deploy on Vercel, the new stack tries to make an agent look like a project directory plus sandbox, workflow, and approval controls rather than a pile of custom glue. |
| GitHub shipped Agent Finder for Copilot, plus new GitHub CLI commands for reading remote repo files and directories without cloning. Agent Finder changelog GitHub CLI changelog | Agent Finder is available on all GitHub Copilot plans; the CLI commands require GitHub CLI v2.95.0 or later. Agent Finder changelog GitHub CLI changelog | Copilot is moving toward discovering approved capabilities on demand instead of stuffing every MCP server, skill, and tool into context up front. |
| JetBrains moved Junie, its AI coding agent, out of beta. The launch post says Junie plans before coding, can drive the IDE debugger, runs async tasks, reviews PRs with project context, and works in IDEs or the terminal. JetBrains Junie announcement | Available in JetBrains IDEs and via Junie CLI; JetBrains says existing JetBrains AI subscribers can use it, and bring-your-own-key/local runtime setups are supported. JetBrains Junie announcement | The pitch is not just code completion. It is agent work that can use the IDE's semantic index, test runner, debugger, and database integrations. |
| Strapi's March-to-June release roundup says v5.47.0 adds a built-in MCP server in beta, using admin API tokens and exposing CRUD-style tools for each content type. Strapi release roundup | The MCP server is opt-in, with server.mcp.enabled false by default; Strapi says admin API tokens are bounded by the owner's permissions and revoked if the owner is deactivated. Strapi release roundup | Teams using Strapi can let agents read or update CMS content through a native interface instead of building a one-off admin bridge. |
| 42Crunch announced an API Security Testing Plugin for GitHub Copilot. It audits OpenAPI specs, flags OWASP API Security Top 10 risks, suggests remediation, validates fixes, and enforces API governance rules. 42Crunch Copilot plugin announcement | Available as a Copilot plugin, according to 42Crunch's announcement. 42Crunch Copilot plugin announcement | If agents are generating API specs and integrations faster, deterministic checks need to sit inside the same loop, not after a human security review queue fills up. |
| Google says Gemini Code Assist IDE extensions stop serving requests for Gemini Code Assist for individuals, Google AI Pro, and Google AI Ultra tiers starting June 18, and the same timeline applies to Gemini CLI. Google Gemini Code Assist deprecation notice | Gemini Code Assist Standard and Enterprise subscriptions are unchanged; consumer users are pointed to Antigravity products and a Gemini CLI migration guide. Google Gemini Code Assist deprecation notice | Check CI scripts, local aliases, IDE login paths, and team onboarding docs if they depend on consumer-tier Gemini CLI or IDE access. |
Agent frameworks are becoming project structure, not just SDK calls
Vercel's eve is the clearest example. The repo describes an agent as a conventional directory:
instructions.md, optional agent.ts, typed tools, skills, channels, and schedules. 1 That sounds mundane, but it is a useful constraint. Files are inspectable, diffable, and easy for both humans and coding agents to traverse.Vercel's larger Ship announcement also wraps eve with production pieces: Agent Stack tooling, Connect for scoped short-lived integration tokens, Vercel Agent for read-only incident investigation by default, and Vercel Services for backend and frontend services under one project preview URL. 2 Treat the adoption numbers in that announcement as vendor-reported, but the direction is concrete: Vercel is designing for code agents that deploy, monitor, and modify software on the same platform.
Loading content card…
For builders, the first test is whether eve's filesystem conventions reduce handoff cost. If an agent's instructions, tools, schedules, and Slack or HTTP channels sit in predictable paths, code review and onboarding should get easier. The tradeoff is lock-in at the workflow layer: even open-source agent structure can become Vercel-shaped once you rely on its sandbox, workflow, deployment, and identity plumbing.
GitHub is putting discovery and repo inspection closer to Copilot
GitHub Agent Finder attacks a different problem: tool overload. Instead of preloading every possible MCP server, skill, canvas, or agent into context, Copilot can search an index of AI resources and return ranked matches from a registry chosen by the user or enterprise. 3 GitHub says the feature implements the open Agentic Resource Discovery specification, developed with Google, GoDaddy, Hugging Face, and Microsoft. 3
The important governance detail is that Agent Finder does not silently install resources. GitHub says enterprises can scope the registry, manage what agents may discover, and keep actual installation under user control. 3 That is the right default for teams already worried about instruction files, MCP servers, and tool permissions spreading faster than review processes.
Loading content card…
The CLI update is smaller but useful.
gh repo read-file and gh repo read-dir let scripts and agents inspect files or directories across public and private repositories without cloning, provided the user has access and is on GitHub CLI v2.95.0 or later. 4 GitHub also changed generated release notes so PRs opened by Copilot cloud agent credit the human who asked Copilot to open the PR alongside @copilot. 5 That is a small attribution fix, but it matters when agent-created PRs start showing up in release histories.IDE-native agents are moving from novelty to workflow surface
JetBrains says Junie is now generally available, with plan-first implementation, debugger control, async remote monitoring, PR review, IDE context, database integration, and CLI access. 6 The strongest part of the pitch is tool use inside the IDE. A coding agent that can set breakpoints, inspect stack frames, run tests, and query project-configured databases has a different failure profile from a chat agent that only edits files and waits for test output.
The benchmark claim needs careful handling. JetBrains quotes Nebius Research Lead Alexander Golubev saying Junie led the latest SWE-Rebench cycle with 61.6% resolved and 72.7% pass@5. 6 That is useful signal, not a blank check. Fresh benchmark tasks help, but your repo's build speed, test quality, dependency graph, and security rules will decide whether Junie is productive or just busier than a human.
The practical experiment is narrow: give Junie one change where the desired behavior is easy to test but implementation touches several files. Require a plan, approve it, then compare the final diff against a human implementation. If the debugger and project index actually shorten the loop, you should see fewer speculative edits and fewer "fix the fix" follow-up commits.
Content, APIs, and security are being wrapped for agents
Strapi's release window is a good example of existing application infrastructure being opened to agents. In v5.47.0, Strapi added a built-in MCP server in beta that authenticates with admin API tokens and exposes content-type tools such as find, create, update, publish, and unpublish. 7 The server is off by default, and Strapi says token permissions are capped by the owner's permissions and re-clamped when roles change. 7
That default matters. A CMS is not just a database with nicer forms. It contains draft content, publishing permissions, locales, media assets, and legal review states. If an agent can create or publish entries, its token model needs to match the editorial permission model. Strapi's first version looks aimed at exactly that boundary.
42Crunch is working the other side of the same problem. Its Copilot plugin checks OpenAPI definitions, identifies OWASP API Security Top 10 risks, gives AI-assisted remediation guidance, validates fixes, and enforces organization rules. 8 The product claim is vendor-side, but the need is real: if agents can draft API specs quickly, reviewers need machine-checkable constraints close to the generation step.
| If this affects your stack | Check first |
|---|---|
| Strapi v5 users experimenting with agents | Whether MCP is enabled, which admin API token owns the session, and whether content publishing permissions are separated from draft-editing permissions. |
| Teams using Copilot for API work | Whether generated OpenAPI specs run through deterministic security checks before code or gateway config is merged. |
| Repos with agent-created PRs | Whether release-note attribution, audit logs, and PR ownership match how your team assigns accountability. |
Google's Gemini CLI shift needs migration checks
Google's deprecation page is short and easy to miss. Starting June 18, Gemini Code Assist IDE extensions stop serving requests for Gemini Code Assist for individuals, Google AI Pro, and Google AI Ultra tiers; Google says the same timeline applies to Gemini CLI and that Login with Google no longer works for those consumer paths. 9 Standard and Enterprise subscriptions are not affected, and consumer users are pointed to Antigravity products. 9
Loading content card…
This is the item most likely to break quietly. Search for Gemini CLI in shell scripts, GitHub Actions, local bootstrap docs, internal wikis, and IDE setup instructions. If a workflow depends on a consumer Google login rather than enterprise access or an API-key-backed path, it needs a migration plan.
What to do with all of this
Do not evaluate these releases as a feature race. Evaluate them as control surfaces.
- Discovery: GitHub Agent Finder and ARD are about letting agents find approved capabilities without permanently loading every possible tool. 3
- Structure: eve and Strapi's MCP server turn agent access into predictable files, permissions, and content operations instead of loose prompts. 17
- Execution: Junie and the new GitHub CLI commands bring agents closer to the repo, debugger, terminal, and review loop. 46
- Guardrails: 42Crunch and Google's deprecation notice are reminders that agent velocity is only useful if security checks and access policies keep up. 89
The next useful question for a team is not "which agent is smartest?" It is simpler: which of these tools can operate inside the permissions, review steps, and failure budgets you already trust?
References
- 1
- 2
- 3
- 4
- 5
- 6Junie: The JetBrains AI Coding Agent Leaves Beta
blog.jetbrains.com
- 7
- 8
- 9Gemini Code Assist consumer accounts
developers.google.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.