KCP manifests — cut ~33% context window usage for CLI tools (Kilo-compatible plugin)

[totto]

If you use Kilo with CLI tools (git, docker, npm, kubectl, etc.) you are likely burning 20–40% of your context window on --help calls and noisy command output.

What this does

opencode-kcp-plugin is a Kilo-compatible plugin that gives the agent structured knowledge of 289 common CLI tools before it executes them. No --help lookup. No parsing noisy output. The agent already knows what flags exist and what patterns to use.

Install

npm install opencode-kcp-plugin

// kilo.json
{
  "plugin": ["opencode-kcp-plugin"]
}

That's it. The plugin is a no-op if no knowledge.yaml is present — but the token savings from manifest-backed CLI context apply regardless.

What the agent sees before running ps aux

[kcp] ps: Report a snapshot of running processes
Key flags:
  aux: All processes, all users, with CPU/memory  → Default
  -ef: All processes, full format with PPID       → Need parent PIDs
  --sort=-%cpu: Sort by CPU descending            → Finding CPU hogs
Prefer:
  ps aux          # Find any process or check what is running
  ps aux | grep <name>  # Find a specific process

Instead of the agent calling ps --help (2,000+ tokens of man page output), it gets a 100-token context block with exactly what it needs.

Numbers (measured across a full agentic coding session)

• 67,352 tokens saved
• 33.7% of a 200K context window recovered
• 73–80% fewer tool calls in explore-heavy sessions
• ps aux output: 30,828 tokens → 652 tokens after noise filter (98% reduction)

Verified compatible with Kilo

Both hooks (experimental.chat.system.transform and tool.execute.after) are present and signature-compatible in @kilocode/plugin. The compiled plugin has zero runtime dependency on OpenCode packages — it works on Kilo out of the box. Tested against @kilocode/plugin v1.2.25.

kilo.json is supported alongside opencode.json — both config filenames work.

What is covered

289 manifests across: git, docker, kubectl, npm/yarn/bun/pnpm, aws/gcloud/az, terraform, python, java, cargo, gradle/mvn, psql/mysql, and 200+ more. Full list: kcp-commands.

Phase B noise filtering is enabled for the high-noise commands: ps aux, docker ps, kubectl get, mvn, terraform, aws, and ~60 others — strips boilerplate and truncates before output reaches the context window.

Links

• Plugin: https://www.npmjs.com/package/opencode-kcp-plugin
• Manifests (Apache 2.0): https://github.com/Cantara/kcp-commands
• KCP spec: https://github.com/Cantara/knowledge-context-protocol

Happy to answer questions or add manifests for tools that are missing from the library.

[decrypt-mm]

Hi. I'm using KiloCode CLI. It sees the plugin loaded, but the agent don't know about this kcp-plugin at all

[totto]

Thanks for the report. Two things to check:

1. Is the hook output actually appearing?

Run a simple command like git status and check whether you see [kcp] git: ... injected before the command runs. If not, the plugin hook is not firing — let me know what your kilo.json looks like and what Kilo version you are on.

2. If the output appears but the agent ignores it

The agent needs to know upfront that injected CLI context will be available. Add a kilo-rules.md (or whatever KiloCode uses for custom instructions) with:

When running CLI commands, structured command hints prefixed with [kcp] may
appear before the command executes. Use them — they contain the right flags
and patterns for the task. Do not run --help when [kcp] context is present.

This is the same mechanism that makes it work in Claude Code: the agent is primed to look for and use the injected context, not just see it as noise.

If neither of those helps, paste what the plugin output looks like when it fires and I will debug further.