Claude Code Digest — Jun 25–Jun 28

Claude Code’s biggest edge this week wasn’t a new model — it was learning that its harness can veto tool calls, fake tool results can be detected, and MCP servers are becoming the default way to wire in real systems.

13,000+ MCP servers

Trending Now

🔥 Harness-Aware Claude Code: Stop Trusting the Surface Prompt
The harness rebuild shows `CLAUDE.md` is layered on top of a hidden base prompt, so your repo instructions are not the whole story. Treat hooks and subagent boundaries as first-class control planes: use them to block unsafe tool calls, enforce formatting, and abort runaway subtrees instead of hoping the model self-polices. 🔥 MCP Is Now the Default Integration Layer — 13,000+ Servers and Counting
The ecosystem crossed 13,000 servers, which means discovery and trust are now the bottleneck, not protocol adoption. If you’re shipping Claude Code workflows, standardize on verified registries like mcp-hub and version your endpoints with `?v=2` so old clients keep working while you iterate. 📈 Production MCP Is Moving Past Demos: OAuth, Filtering, Rate Limits
The Snowflake and PLM examples show the bar has shifted from “can it connect?” to “can it survive enterprise use?” Build MCP servers with OAuth 2.0, schema filtering, and rate limiting from day one, or you’ll end up with a prototype that security teams block on sight.

Best Practices

Add a fake-result detector hook to catch hallucinated tool output
Without this: Claude can claim `` blocks or `written: N bytes` outputs that never happened. With a `zen_stop_hook` grep check on every turn, you can flag fabricated tool results immediately and force a real tool re-run. Version MCP servers with `?v=2` instead of breaking schemas
Without this: every schema change forces a client update and breaks older Claude Code setups. With query-param versioning, old and new clients coexist against the same server URL with zero client-side code changes. Use a body-vs-allowlist cross-checker to verify agent behavior
Without this: your 317+ agents can drift from declared permissions and silently do more than policy allows. With a versioned cross-checker, you can automatically prove the body matches the allowlist after every tightening rule.

Tools & MCP

mcp-hub — Verified MCP search/install — cuts discovery time and avoids broken servers in a 13,000+ server ecosystem FastMCP + pgvector server — Turns hardcoded RAG into a reusable vector-search MCP server — lets any LLM client query your database without custom glue Snowflake MCP server — Enterprise-ready data access with OAuth 2.0, schema filtering, and rate limiting — safer than exposing raw warehouse access

Multi-Agent Patterns

Abort-Tree Subagents
Treat subagents like a tree with explicit abort paths, so a bad branch can be killed without nuking the parent task. This is the cleanest way to keep parallel exploration from turning into runaway token burn. Queued Prompts + Plugin Marketplace
Cline v4.0.0’s queued prompts and marketplace point to a more composable agent stack: queue work, load capabilities as plugins, and reduce the need for one giant always-on context. Parallel GitHub Actions with `background`/`wait`/`cancel`
Use the new workflow keywords to fan out checks in parallel, then cancel stale branches as soon as one fails. That’s the CI pattern Claude Code teams should mirror for faster agent validation loops.

Community Requests

• Native MCP server discovery with trust signals, install health, and deprecation warnings baked into Claude Code
• Built-in tool-result integrity checks so fake `` blocks and fabricated file-write claims are caught automatically
• First-class harness controls for subagent abort trees, hook vetoes, and per-tool policy debugging