holstered

Your agent has 500+ skills holstered. It draws none. holstered draws the right one.

Fastest draw in the terminal: one skill, or none, in about a second.

English · 简体中文

You type "make me a 5 slide deck about our Q3 results":

| | Without holstered | With holstered | |---|---|---| | What the agent knows about your PowerPoint skill | One line, lost among 581 others | The full skill, loaded before it starts | | What happens next | It may write the deck from scratch and ignore your skill | It follows your skill's instructions |

You type "thanks, that fixed it": no skill fits, so holstered adds nothing.

Agents with hundreds of skills see only their names, so the one that fits the task often goes unused. holstered is a prompt hook: on every user prompt it shortlists skills with BM25, asks a decision model to pick one (or none), and injects that skill's SKILL.md into the model's context for the turn. A prompt that spans several tasks also gets up to 2 runner-up skills over the runner-up threshold.

https://github.com/user-attachments/assets/b6de382b-8ca3-4fe0-9baa-0039342dd3b0

The decision model is your choice: Jev, hosted on OpenRouter, or Kev, an open model you run locally so nothing leaves your machine. See Choose a decision model.

One Rust binary serves every agent through polyhook, which translates each agent's hook payload and response format.

How it works

Nothing is injected, and the prompt goes through untouched, when the decision model answers none, names a skill it was not offered, fails or times out (8s by default), or none is configured. holstered never blocks a prompt.

An injected SKILL.md is capped at 8,000 bytes, since some agents cut hook output past ~10KB. A longer skill is cut there and ends with a pointer to the file for the rest, which the model may not follow. Keep SKILL.md under 8KB (about 100 lines of prose; check with wc -c) and move detail into files it references.

Why BM25 + a decision model

On a 38-prompt labeled set over a 581-skill library (32 prompts with a correct skill, 6 with none):

The holstered rows are the release binary run end to end against live Jev (median 915 ms per prompt) and a local Kev-4B on an Apple Silicon Mac (median 1.7s). One of its three misses picked a pptx skill for a slide-deck prompt the labels credited only to another skill. Small set, single labeler: read it as a direction, not a guarantee.

Supported agents

Agent support comes from polyhook; see its Supported Tools for the full list and which agents can take context injection.

| Agent | Hook | Injection | |---|---|---| | Claude Code | UserPromptSubmit | ✅ | | Codex | UserPromptSubmit | ✅ | | Gemini CLI | BeforeAgent | ✅ | | Hermes Agent | pre_llm_call | ✅ | | Cline | UserPromptSubmit | ✅ | | Cursor, Windsurf, Amp | — | ❌ their prompt hooks can only allow or block, so there is nothing to register |

Install

Pick one:

# npm (prebuilt binary, no Rust toolchain)
npm install -g holstered

# Go (downloads the prebuilt binary on first run)
go install github.com/tupe12334/holstered/go/cmd/holstered@latest
# or pin it in a Go project (Go 1.24+): go get -tool github.com/tupe12334/holstered/go/cmd/holstered

# crates.io
cargo install holstered

# Homebrew (builds from source)
brew install tupe12334/tap/holstered

# latest main from GitHub
cargo install --git https://github.com/tupe12334/holstered

Then choose a decision model and add holstered to each agent you use:

Claude Code — install the plugin (send the two commands as separate prompts):

/plugin marketplace add tupe12334/holstered

/plugin install holstered@holstered

Codex — install the plugin, then open /hooks in codex and trust its hook:

codex plugin marketplace add tupe12334/holstered
codex plugin add holstered@holstered

Gemini CLI — install the extension:

gemini extensions install https://github.com/tupe12334/holstered

Hermes Agent — install the plugin, then restart Hermes:

hermes plugins install tupe12334/holstered#plugins/hermes --enable

Cline — hooks are executables named after the event

mkdir -p ~/Documents/Cline/Hooks
ln -s "$(command -v holstered)" ~/Documents/Cline/Hooks/UserPromptSubmit

Choose a decision model

Set one of these in the environment your agent starts from. If both are set, Kev wins.

| | Jev (hosted) | Kev (local) | |---|---|---| | Set | OPENROUTER_API_KEY | HOLSTERED_KEV_URL | | Runs | OpenRouter Decisions API | Kev server on your machine | | Privacy | prompt (first 500 + last 1,500 chars) and shortlisted skill descriptions go to OpenRouter | nothing leaves the machine | | Latency per prompt | ~1s | ~1.7–3.5s on an Apple Silicon Mac | | Accuracy on the set above | 29/32 | 26/32 |

Kev setup and first-request timeout: CONFIGURATION.md.

Configuration

Environment variables, the data sent to OpenRouter, and running fully offline with a local Kev model: see CONFIGURATION.md.

Contributing

See CONTRIBUTING.md.

License

MIT