MCP Bridge Code Mode Reduces Context Window by 98%

Large APIs with many tools pose the next challenge in agents blowing out context windows just trying to understand what tools to use. Ultimately leading to poor performance and high costs through token consumption. Bring the systems you have today for agentic workflows without compromise. #mcp #agentic #devtools #ai #techdebt #modernization

Tool definitions eat context windows. ~250 tokens per definition × 100+ operations on a real enterprise API = 50,000 tokens of schema before your agent has done anything. 🤯 Today we shipped Code Mode on MCP Bridge. Instead of loading every tool definition, the agent gets 3 meta-tools: → search_tools — describe what you need, find the right tool → get_tool_schema — pull a specific tool's schema on demand → execute_code — run multi-step tool calls in a sandboxed JS runtime Result on real enterprise API surfaces: ~98% context window reduction. The model only sees the tools it picks, only when it needs them. Bonus: loops and conditionals live in execute_code. 800 customer lookups ≠ 800 inference round trips. @Cloudflare and @Anthropic have both written about Code Mode. What's new: it's built into a platform that already translates REST, GraphQL, SOAP, and gRPC into MCP. Code Mode without giving up multi-protocol coverage. Benchmarks + early access in comments. Launching on Product Hunt this Thursday 🚀 #AppFactor #MCP

The reason your MCP agent picks the wrong tool isn’t your code. It’s three sentences. When you give Claude or Cursor 20 MCP tools, the model doesn’t run them to decide which to use. It reads their descriptions and parameter schemas. That’s the entire selection process. I spent the last few months building MCPForge - a hosted platform that generates MCP servers from OpenAPI and GraphQL specs. I expected the engineering to be the hard part. GraphQL schema parsing in particular took longer than it should have. But those were finite problems - solve them once and move on. The hard problem turned out to be tool descriptions. Two failure modes kept coming back, both worse with smaller models (GPT-4o-mini was the consistent offender): 1. Tool selection hallucination. The model would call the wrong endpoint with confidence. The fix wasn’t smarter prompting - it was tightening the description so the boundary between similar tools was unambiguous. 2. Input field mapping (the worse one). The model would call the right tool but pass garbage arguments - wrong format, wrong field, hallucinated values. “User ID” as a parameter description does nothing. “The customer’s stable UUID from the auth token, not the display ID in the UI” does everything. Three sentences. That’s the gap between an MCP integration that works in demos and one that holds up in production. And the MCP discourse is almost entirely about generating servers (where Gram, Stainless, FastMCP, and the open-source generators all converge), not steering tool selection. What I’m exploring next: can you systematically improve descriptions instead of hand-editing them? Watch where the model fails, propose better wording, test it. Nobody seems to be building this layer yet. I suspect it becomes the dominant problem in MCP within a year. If you’ve shipped MCP work in production, I’d be curious what you’re seeing. #MCP #AIInfrastructure #LLM #AIEngineering #BuildInPublic

Built an MCP server this weekend. Auth, idempotency, observability, schema versioning — the parts most MCP demos skip. Repo (MIT, TypeScript): https://lnkd.in/ghasGTzm Full write-up: https://lnkd.in/g9eD58rb The 7 things between an MCP demo and an MCP server you'd hand to a paying user: 1. Bearer auth + scope claim per tool (not per server) 2. Versioned tool registry (so signature changes don't break in-flight clients) 3. Idempotency keys on every mutating tool call 4. Structured logging — input / output / latency / cost on every call 5. Deprecate-then-remove for tool removals (not just delete) 6. Cost instrumentation per tool, not per request 7. Schema validation that survives Anthropic SDK updates Most public MCP examples stop at step zero. This is what "production-grade" means once somebody pays you. Concrete number from my build: 80% of the cost was in one tool — a doc-search reranker on Sonnet 4.6. Caching reranks for repeat queries took it from £400/month to £8/month at the same usage. Per-tool cost is the only way you find that. ~600 LOC on top of the Anthropic MCP SDK, 47 vitest cases, stdio + HTTP entry points. Config snippets for Claude Desktop and Cursor in /examples. Fork it, copy the patterns you need, or use it as the start of your own. Reply with what you'd want an MCP server to do for you next.

My solution to the problem became the problem. Here's how that led me to build CMDR. A couple years ago I was a frontend engineer being cross-trained into backend work. I was excited, it felt like leveling up. Then I saw what testing my code actually required. Serverless, Optimus, AWS. Flags, namespacing, variables that changed from developer to developer and environment to environment. There was no chance I was remembering any of this. So I did what felt natural. I wrote a bash script to wrap all the complexity and prompt for the variables that would change. Simple. Clean. One command instead of five. People loved it. The whole team started using it. So when I got onboarded into a second project, I wrote more scripts. Then a third. More scripts. I now had dozens of them and couldn't keep track of the names, let alone the commands. The solution had become the problem. I sat down one night and stared at all of them laid out together. And it hit me, these are just structured data. Steps, variables, flags. They could all live in a single config file and build themselves. Shareable. Version-controlled. Self-describing. So I started building that. That's CMDR. A platform for teams to create, share, and run CLI workflows. So the knowledge lives in the tool, not in someone's bash history or a Slack thread from six months ago. Still early days. But the problem is real. I lived it.

RunMQ and its NestJS module are now at v2.0.0 including massive reliability improvements, with a fresh documentation page 🥳 This release doesn't introduce any new major features; it's about closing every "what if?" I still had it in the back of my head about running RabbitMQ-backed queues in production. Publishers now use Publisher Confirms instead of fire-and-forget; the trade-off is not easy, throughput dropped from ~200k to ~70k msg/s, but every publish is now broker-acknowledged, which improves resilience and reliability (still about 1.3× higher than BullMQ) A few other things I'm excited about: - Channels auto-re-subscribe when they drop, consumers will come back automatically instead of going down silently due to Channel timeouts or any other issues. - Internally, publishing to DLQs is done in two steps; this fits with the at-least-once delivery guarantees, but on RabbitMQ >= 4.x, we can rely on the Modified Outcome feature for atomic routing, this's coming next! Stay tuned, and check the docs out @ https://lnkd.in/esewNj6C 🚀

Most coverage of K2.6 missed the more strategic launch. Kimi Code, the CLI agent that ships with K2.6 as its default backend. If I had to describe what Moonshot did with Kimi Code in 1 word: 'unbundling.' It's a clean unbundling of the 3 things Anthropic locked together with Claude Code. The model. The CLI. The tool ecosystem. Each one's now portable. As of early May 2026, Kimi Code's at 6,400+ stars on GitHub. It ships under Apache 2.0 (the CLI itself, separate from the model's Modified MIT). And it does 3 things that change the calculation: 1. Shell-aware. Ctrl-X toggles into bash mode so you run shell commands inline without leaving the agent. Same flow as Claude Code's bash integration. 2. MCP servers from Claude Code work in Kimi Code without modification. The Anthropic ecosystem of community-built tools just became portable. Every MCP server you wired into Claude Code runs unchanged. 3. Implements Agent Client Protocol. Zed, JetBrains, and other ACP-compatible editors talk to Kimi Code as a backend agent. The IDE integration story isn't 'wait for someone to build a plugin.' It already works. The strategic move's clean and brutal. Take Claude Code's interaction model. Make it open-source. Plug into the existing tool ecosystem. Charge a fraction. Anthropic now has a competitor that benefits from every MCP server its own community builds. That's a structural problem, not a marketing one. If you're a Claude Code user paying $20+ a month for Pro and another $50+ in API spend, run this experiment this week: 1. Install Kimi Code locally 2. Point your existing MCP servers at it 3. Run your standard daily workflow on K2.6 for 5 days 4. Compare cost, latency, and completion quality Come back and tell me what you find. ♻️ Share this to inspire someone. ➕ Follow me, Ankur Kesharwani, to stay in touch.

Crossview - Crossplane Dashboard A modern React-based dashboard for managing and monitoring Crossplane resources in Kubernetes. Visualize, search, and manage your infrastructure-as-code with ease. https://lnkd.in/garEJAqz

I gave a talk for years on Configuration as Code. The story behind it was a project where we had to modernize a system buried under layers of config accumulated over time, with a mandate to get to zero downtime. The only way through was to stop treating configs as a separate species from code, so we versioned them, tested them, deployed them through the same pipeline as everything else. All of this came back to me reading Cloudflare's writeup on their late-2025 outages (yes, Cloudflare again, they've been putting out a lot of interesting stuff lately), both triggered by config changes rather than code, and the engineering effort that followed to give configuration the same rollout discipline as code. The other thing in that writeup worth pointing at is what they call the Codex: incident learnings turned into versioned rules, enforced through automated code review, so that the next merge request that would have caused the same outage gets rejected before it ships. Configuration as Code ==> Lessons as Code. Same idea applied one layer up. https://lnkd.in/dEerhHyf

I got tired of treating deployments like a black box. So I decided to learn system design to understand how the systems work and create a mini Vercel platform Introducing ShipStack, a production-ready deployment platform. 🚀 What it does: Full deployment pipeline (queued → running lifecycle) Auto Docker builds based on project type Smart port allocation + subdomain routing Real-time logs via SSE Self-healing containers 🔐 Built with security in mind: * JWT auth + refresh rotation * Redis-backed token blacklisting * TOTP 2FA + OAuth (Google/GitHub) * AES-256 encryption for sensitive data ⚙️ Under the hood: * Nginx reverse proxy * BullMQ + Redis for async deployment queues * Health checks + auto-restarts 💡 Biggest takeaway: This project forced me to think in systems not just features. If you’ve ever relied on deployment platforms without understanding them try building one. It’s worth it.