Reduce MCP token usage

MCP servers are the fastest way to blow through a context window. Every tool description loads up front, and every tool response comes back raw - a full kubectl get pods, a whole docker ps table, a page of JSON. banish compacts that output at the tool layer, before it ever reaches Claude Code, Cursor, or any MCP agent.

Why MCP servers eat your context window

Two things spend tokens before you do any real work:

• Tool descriptions load on the first message. A handful of MCP servers can add ten to twenty thousand tokens of schema before you type anything.
• Tool responses return verbatim. A single verbose command can push tens of thousands of tokens of raw text back into the window.

The result is the common complaint: multiple MCP servers running, and half the context window gone before the agent starts.

When tool output is too large

Naive fixes make it worse. Blindly truncating a response to a fixed byte count throws away the part you needed and leaves the noise. banish compacts instead of truncating: it runs the real command, pipes the output through a matching filter, and returns compact text that keeps the signal.

If a filter does not match or fails, you get the raw output back. banish never swallows data.

Compacting MCP responses with banish

Run the server and every verb in your extensions becomes a compact MCP tool:

banish serve

With the default extensions, agents get 45-plus tools out of the box, each a compact wrapper around a command you already run:

banish_gs      git status, compact
banish_dps     docker ps, compact
banish_kpods   kubectl get pods, compact
...

The proxy and the server share the same filters, so you get the same compact output whether your agent shells out or calls a tool. This is an MCP-native fix - the compaction happens at the tool boundary, not through a fragile wrapper around one shell.

Set it up in one command

Register the server with any MCP-capable agent:

banish init mcp

For Cursor, run banish init cursor. See set up an agent for the full list, or read how the MCP server exposes your verbs as tools.