Apideck CLI: A Low-Context Alternative to MCP for AI Agents
MCP's Context Window Problem
The article describes a specific scenario: connecting GitHub, Slack, and Sentry via MCP servers (about 40 tools total) consumes 55,000+ tokens before the agent processes a single user message. That's over a quarter of Claude's 200k limit. Each MCP tool definition costs 550–1,400 tokens for its name, description, JSON schema, field descriptions, enums, and system instructions.
One team reported three MCP servers consuming 143,000 of 200,000 tokens (72% of the context window), leaving only 57,000 tokens for actual conversation, retrieved documents, reasoning, and response.
David Zhang (@dzhng), building Duet, described removing MCP integrations entirely due to a "trilemma": load everything up front and lose working memory, limit integrations so the agent can only talk to a few services, or build dynamic tool loading which adds latency and middleware complexity.
Benchmark Results
A benchmark by Scalekit ran 75 head-to-head comparisons (same model: Claude Sonnet 4, same tasks, same prompts) and found:
- MCP costs 4 to 32× more tokens than CLI for identical operations
- Checking a repo's language consumed 1,365 tokens via CLI vs 44,026 via MCP
- The overhead is almost entirely schema: 43 tool definitions injected into every conversation, of which the agent uses one or two
Apideck CLI Approach
The Apideck CLI uses an ~80-token agent prompt that replaces tens of thousands of tokens of schema. It features progressive disclosure via --help and structural safety baked into the binary. Any agent that can run shell commands can use it without requiring protocol support.
Industry Responses to Context Bloat
The article identifies three industry approaches:
- MCP with compression tricks: Compress schemas, use tool search to load definitions on demand, or build middleware that slices OpenAPI specs into smaller chunks. This works for small, well-defined interactions but adds infrastructure needs (tool registry, search logic, caching, routing).
- CLI-first interfaces: The approach taken by Apideck CLI.
- Agent-native protocols: Mentioned but not detailed in the provided source text.
📖 Read the full source: HN AI Agents
👀 See Also
Developer shares hybrid AI coding workflow: Claude for planning, local models for execution
A developer built a pipeline using Claude 3.5 Sonnet for task planning and local Qwen2.5-Coder models via Ollama for code generation, achieving 85% token reduction compared to using Claude alone.
Building a Persistent AI Knowledge Infrastructure with OpenClaw
A developer built 'Brain'—a central knowledge service with local RAG, multi-agent coordination, and a typed plugin system—to solve the statelessness problem in AI setups. The system runs entirely on local hardware using Ollama, Postgres, MongoDB, Qdrant, and Memgraph.
Kula: Self-contained Linux server monitoring with zero dependencies
Kula is a lightweight Linux server monitoring tool that runs as a single binary with no external dependencies or databases. It collects system metrics every second from /proc and /sys, stores them in a built-in tiered ring-buffer, and provides both web dashboard and terminal TUI interfaces.
agent-recall: Local SQLite MCP for Persistent Claude Code Memory
agent-recall is an MCP server that gives Claude Code persistent memory across sessions using a local SQLite file. It provides 9 MCP tools for saving entities, relationships, and observations, with LLM-summarized briefings at session start instead of raw data dumps.