Claude Code is basically my second employee.
I have tried 100s of MCPs by now.
Most of them are bloated and CLI tools are much much better.
In fact, most MCP integrations are burning tokens for nothing.
First let’s talk about:
The token cost (which nobody talks about)
And here is the part that stings.
Loading the schema for a GitHub MCP server with 26 tools costs around 4,200 tokens before the agent does anything at all.
Filesystem MCP with 14 tools costs about 3,400.
Stack five of these wrapper servers and you have spent 15 to 20 thousand tokens just showing your agent a menu of tools it will mostly never call.
And yet founders are like: let’s use an MCP.
That is context you are paying for on every single message.
Not once.
Every turn.
Secondly,
The actual rule people should converge on
Think about the tasks you actually do, then pick tools for those tasks.
Not the other way around.
Basically,
Do not install an MCP server because it exists.
If a tool already has a good CLI, use the CLI.
Claude and Codex both have enough training data on git, gh, and most common CLIs that they do not need a wrapper around them.
I am doing exactly this.
My GitHub workflow, PRs, issues, releases, all of it now runs through gh in the terminal instead of a GitHub MCP.
Same result, none of the schema tax.
Now, just think: do you need an mcp anywhere?
Maybe.
So,
What actually earned its keep (according to my stack)
These are genuinely worth the token cost, and they map closely to what I actually use:
Browser automation. Playwright and chrome-devtools MCP.
The reason is simple: a growing number of sites now actively block AI crawlers, including Reddit, which is exactly why I built my own Splitwise MCP server and why several people are building custom Reddit MCPs just to get research data at all. If a site blocks your agent from reading it any other way, a browser-based MCP is not overhead, it is the only path in. And they do have a CLI, but this one feels easy and is still cost effective.
Your own internal tools.
This is probably the strongest theme. The MCP servers people actually valued are the ones they build themselves to talk to their own apps, internal dashboards, or proprietary data. Generic public MCPs mostly got dismissed as noise. Custom ones solving a real, narrow problem earned trust.
Read-only data access with hard guardrails.
A read-only Postgres or MySQL connection, sometimes through a dedicated database user with zero write permissions. That is the difference between an MCP that saves you time and one that becomes a liability the day your agent hallucinates a destructive query.
Anything genuinely behind auth with no CLI alternative.
Jira, Sentry, Figma.
These come up as fair use cases specifically because there is no equivalent command line tool that does the same job as cleanly.
So, in short:
Cutting from your own setup
- GitHub work moves entirely to gh CLI. No more GitHub MCP loaded by default.
- Filesystem MCP gets dropped. Claude Code already has native file tools and does not need a wrapper duplicating that.
- Any MCP not actively using in a session gets disconnected before you start, not left running in the background eating tokens
If you are running a lean operation like I am, where every hour of build time actually matters, the MCP sprawl trend is a trap.
It feels productive to connect ten integrations.
It is not.
What actually moves your product forward is a short list of tools that do real work your agent cannot already do through a shell, plus whatever custom connector you build for your own stack.
techaiguild.aibucket.org
In case we are meeting for the first time, come over here, it’ll be worth the roller coaster of articles that are gonna come up in the next few weeks.