TL;DR: MCP eats context, has low reliability, and overlaps with existing CLI/API.
<aside>💡
Reference: MCP is dead. Long live the CLI
After reading the above article, we ran the experiments on our actual stack. This document covers the original argument, additional research, and our measurements.
</aside>
<aside>📌
Update: Since these measurements were taken, Claude Code has rolled out Tool Search with Deferred Loading, which loads MCP tool schemas on-demand and reduces context usage by 85%+. The context bloat described in Problem 1 is largely addressed for users on current Claude Code versions. The performance, debugging, and architectural arguments below still apply.
</aside>
MCP (Model Context Protocol) connects LLMs to external tools (GitHub, Linear, Notion, Slack, etc.).
Since its launch in late 2024, it's been called "the USB-C of the AI ecosystem." But developers actually using it day-to-day are starting to think differently.
TL;DR: MCP eats context, has low reliability, and overlaps with existing CLI/API.
The context window is the LLM's desk. When you connect MCP servers, tool definitions alone take up a significant chunk of that desk.
Restaurant analogy:
We extracted and measured the actual tool definitions from the MCP servers connected in our environment. With all 4 servers connected, 10.5% of the context window is consumed by tool definitions alone.
MCP ServerToolsEstimated CharsEstimated TokensLinear42~51,229~12,807Notion14~16,156~4,039Slack12~15,168~3,792Postgres9~1,755~438Total77~84,308~21,077
ModelContext WindowUsage by Tool DefinitionsClaude (200K)200,000 tokens10.5%GPT-4o (128K)128,000 tokens16.5%
Linear alone accounts for over 12,800 tokens. That's 42 tool definitions always loaded, even if you only ever use get_issue and save_issue.
ToolChars~Tokenslinear/save_issue2,479~619slack/search_public1,614~403linear/list_issues1,588~397notion/fetch1,379~344slack/send_message1,248~312
IssueDetailInit failure, repeated re-authRequires starting and maintaining a separate processSlower AI responsesExternal server round-trip on every tool callMid-session tool deathMCP server process crashesOpaque permissionsUnclear what permissions each tool actually has
Performance is a known issue. The author of the original article benchmarked Jira MCP against its REST API directly and found MCP was 3x slower per call, and 9.4x slower on first call including initialization. This isn't Jira-specific, it's architectural: every MCP server adds a process layer between the LLM and the underlying API. The same overhead applies to the Linear, Notion, and Slack servers in our stack.
AspectCLI / APIMCPHuman-machine paritySame commands for humans and LLMsOnly exists inside LLM conversationsComposabilityPipes, jq, grep freely combinableLocked to server return formatDebuggingReproduce immediately in terminalOnly reproducible inside conversation contextTraining dataAlready learned from man pages, StackOverflowRequires separate tool definitionsInstall costMostly already installedServer setup, auth, process management needed
How many tokens does it cost to look up the same Linear issue?
MCP consumes ~65x more tokens than the CLI approach.
[ CLI approach: ~200 tokens ]
curl -s -H "Authorization: Bearer $LINEAR_TOKEN" \\
-H "Content-Type: application/json" \\
-d '{"query":"{ issue(id: \\"ISSUE-ID\\") { title state { name } assignee { name } } }"}' \\
<https://api.linear.app/graphql>
-> Prompt (curl command): ~50 tokens
-> Response: ~150 tokens
[ MCP approach: ~12,957 tokens ]
-> Tool definitions (always loaded): ~12,807 tokens (42 tools)
-> Tool call + response: ~150 tokens
Provide CLI -> API -> docs, in that order. LLMs already learned from man pages and StackOverflow.
Using existing CLI directly:
If MCP is "spreading all menus on the table upfront", Skills is "asking the librarian for only the book you need".
AspectMCPSkillsLoading timeAll tool definitions loaded on connectOnly loaded when neededContext consumptionAlways occupiedOnly when in useScalabilityContext pressure grows with each serverNot proportional to skill count
The key is embedding CLI usage instructions inside Skills. Combined with Alternative 1's CLI-first strategy, this is most efficient. For example, a Linear skill:
# Linear Issue Lookup Skill
- Linear API: <https://api.linear.app/graphql>
- Auth: Bearer Token ($LINEAR_TOKEN env var)
- Get issue: curl -s -H "Authorization: Bearer $LINEAR_TOKEN" -H "Content-Type: application/json" -d '{"query":"{ issue(id: \\"ISSUE-ID\\") { title state { name } assignee { name } } }"}' <https://api.linear.app/graphql>
- Search issues (GraphQL): adjust the query field for JQL-like filtering
- Results are JSON, parse with jq
This way, the LLM only loads the above into context when the skill is invoked. No need to carry 42 tool definitions at all times. Just the CLI commands it needs.
Not entirely. MCP is still valid when:
Short answer: it depends.
DBs are just query execution at the end of the day. LLMs already know SQL and MongoDB queries well. Put DB info and CLI usage in a skill, and it works fine without MCP. Just give it the schema and it writes the queries.
# Postgres Skill
- Host: postgres://localhost:5432/myapp
- Tables: users (id, name, email), orders (id, user_id, status)
- CLI: psql -h localhost -d myapp -c "SELECT * FROM users WHERE ..."
However, MCP has advantages for databases:
DROP TABLE.ScenarioRecommendationWhyLocal dev / personal DBSkills + CLILight and fast. Easy to recover from mistakes.Production DB / shared teamMCPSafety guardrails are essential. Query validation and access control at the server level.
But for most developer workflows, MCP is over-engineering.
These days, every SaaS landing page has "MCP supported" in the feature list. Whether the MCP server is stable or how much context it eats doesn't matter - the goal is checking the "we do MCP too" box. Same pattern as "AI-powered" and "blockchain-based" marketing from years past. When users actually connect, they get dozens of tool definitions loaded, initialization failures, and mid-session crashes.
At Quandri we use all three approaches side by side, picking what fits each service:
gh, psql, aws). Zero context cost, full flexibility, debugs straight in the terminal.We don't force one path. If a CLI already exists and authenticates locally, that's usually the lightest option. If a service has no CLI or we need uniform auth across the team, MCP earns its keep.
Teaching well matters more than connecting everything.
For us, replacing MCP servers with Skills that wrap existing CLIs freed up ~21K tokens of context, removed init failures from our daily workflow, and kept debugging in the terminal where it belongs.
Load only the tools you need, only when you need them, with CLI instructions baked in. MCP might evolve to solve these problems, but right now, Skills win.
<aside>📊
Measurement methodology: Tool definition sizes were measured by extracting the JSON schema of each tool (name + description + parameters) from actually loaded MCP servers in our Claude Code environment. Token estimates use the ~4 chars/token heuristic. Full server estimates are extrapolated from sampled tool averages.
</aside>