Lockkeeper

The skill router and prompt-injection firewall for AI coding agents

Give Claude Code, Codex, Cursor and other AI agents the few skills, MCP servers and tools that fit each task, instead of all of them.

Smaller context window, better tool choices, and no unvetted skill instructions reaching your agent.

Quickstart · Ways to use it · Benchmark · Firewall · FAQ · Docs

---

What is Lockkeeper?

AI coding agents get better with skills (SKILL.md files), MCP servers, plugins and tools. But every one you install adds to what the agent has to read and choose from. With hundreds installed, your context window fills up before work starts, and the agent often picks the wrong skill or none at all.

Lockkeeper is a local skill router. It indexes everything installed across all your agents, and for each task it hands the agent a small, complementary set, up to 10 capabilities by default (you choose the size), with the exact file to read for each. Before anything reaches your agent, its built-in firewall can check skills and live tool calls for prompt injection.

$ lockkeeper route --runtime claude "migrate the auth module to the new token API"

[primary] skill: api-migration [context] mcp: context7 [integration] tool: mcpcontext7query_docs [verification] agent: code-reviewer [support] skill: python-patterns context savings: loaded 6 of 7,540 eligible capabilities (7,534 kept out of context)

Why developers use Lockkeeper

• 🎯 Better skill choices. On a public benchmark of real agent tasks, Lockkeeper ranks a correct skill first 65% of the time among 26,000 real skills (up from 35%) and 55% among 79,000, no model required. See the benchmark.
• 📉 A context window that stays small. With 58,018 capabilities in the library, a routed task still carries a median of about 8,700 tokens of skills instead of about 77.5 million. Adding skills to the library doesn't grow your prompt.
• 🛡 Safer skills and plugins. Scan any skill, plugin or MCP config for hidden instructions and data exfiltration before your agent reads it, and block hostile tool calls live.
• 🔌 Works where you already work. Automatic routing in Claude Code, an MCP server for Codex, Cursor, Windsurf, Cline and other clients, and a CLI for everything else.
• 🔒 Local, private and dependency-free. Pure Python standard library. No GPU, API key or cloud service needed. Telemetry is off unless you say yes.

Quickstart

1. Install from PyPI (Python 3.11+, macOS, Linux and Windows):

pipx install lockkeeper     # or: pip install lockkeeper  ·  uv tool install lockkeeper

lockkeeper init # finds every AI agent on this machine and connects it

Only want the router skill? npx skills add Hannay001/lockkeeper installs it for any agent (it needs the lockkeeper command too). Prefer not to use a terminal? Paste the prompt in PROMPT.md into the AI agent you already use; it installs and configures Lockkeeper for you. Working from source? git clone https://github.com/Hannay001/lockkeeper.git && cd lockkeeper && ./install.sh

2. Index what you have installed:

lockkeeper rebuild    # indexes every skill, agent, command, MCP server and plugin it finds

lockkeeper doctor # shows each agent found and how many skills it has

3. Route a task:

lockkeeper route "write unit tests for a python data pipeline"

Then pick how your agent should use it, below.

Four ways to use Lockkeeper

1. Route every prompt automatically (Claude Code)

Install the Claude Code plugin (after pipx install lockkeeper). Inside Claude Code:

/plugin marketplace add Hannay001/lockkeeper

/plugin install lockkeeper@lockkeeper

It adds the routing hook, the MCP server and the router skill in one step. Prefer settings files? lockkeeper hooks install claude adds just the hook (use one or the other, not both).

Every prompt you send now reaches Claude Code with a short note naming the installed skills that fit it and the exact files to read. Slash commands and short replies like "thanks" pass through untouched, and the hook never blocks a prompt. Undo with lockkeeper hooks remove claude.

Then shrink the list Claude Code loads into every session:

lockkeeper library move            # shows the plan: which skills move, how many tokens it saves

lockkeeper library move --apply # moves them to ~/.agents/library; the hook still finds them lockkeeper library restore --apply # puts them all back

Claude Code puts the name and description of every skill in ~/.claude/skills into each session. Library mode moves them to a folder Lockkeeper indexes but Claude Code doesn't load, so only the skills a prompt needs reach the context. Keep favorites where they are with --keep NAME.

2. As an MCP server (Codex, Cursor, Windsurf, Cline and any MCP client)

lockkeeper mcp gives your agent three tools, route, search and audit, and keeps the index loaded between calls so answers are fast.

claude mcp add lockkeeper -- lockkeeper mcp          # Claude Code

# Codex: ~/.codex/config.toml

[mcp_servers.lockkeeper] command = "lockkeeper" args = ["mcp", "--runtime", "codex"]

{ "mcpServers": { "lockkeeper": { "command": "lockkeeper", "args": ["mcp"] } } }

The JSON form works for Cursor (~/.cursor/mcp.json), Windsurf, Cline and most other clients. Lockkeeper is also listed in the official MCP Registry (MCP Registry name: mcp-name: io.github.Hannay001/lockkeeper).

3. From the command line and scripts

lockkeeper route --runtime codex "add rate limiting to a REST endpoint"

lockkeeper search "pdf tables" lockkeeper route --json --stdin < task.txt # whole prompts, machine-readable output

4. As a firewall for skills and plugins

lockkeeper audit ~/Downloads/some-skill --recursive --strict   # exit 2 = hostile

lockkeeper hooks install claude --firewall # block hostile tool calls live

Supported agents

Lockkeeper reads the formats you already use: SKILL.md Agent Skills, agents and commands in Markdown, plugin manifests, and MCP server configs. The installer also detects agent tools it doesn't know by name.

Proven on a public benchmark

Routing claims should be measurable. Lockkeeper is tested against SkillRouter Eval Core, the public benchmark from the SkillRouter paper (arXiv:2603.22455): 75 real agent tasks with known correct skills, hidden among real SKILL.md files from public repositories, including 780 deliberately misleading look-alikes.

On the full pool, Lockkeeper's standard-library ranker scores between the paper's general-purpose embedding models (Qwen3-Embedding-0.6B at 53.3%, Gemini embedding at 56.0%) and roughly double its BM25 keyword baseline (28.0%), without loading a model. Methods, per-change results and caveats: [docs/BENCHMARK.md](https://github.com/Hannay001/lockkeeper/blob/main/docs/BENCHMARK.md).

Reproduce it yourself (downloads the ~400 MB dataset once):

python3 scripts/bench_routing.py prepare --home /tmp/lk-bench --size 26000

python3 scripts/bench_routing.py run --home /tmp/lk-bench

Your prompt stays flat as your library grows. On the 79,141-skill benchmark pool (about 157M tokens of skill text), six everyday tasks each routed to 10 capabilities: a median of about 16,000 tokens even if the agent reads every one, over 99.98% kept out of context. The 26,000-skill pool gave about the same (17,600). Reproduce with python3 scripts/bench_context_savings.py.

How it works

flowchart LR

T["Your task or prompt"] --> S["Rank every installed capability names, descriptions, body keywords, rare words weighted higher"] S --> P["Apply your policy deny lists, required roles"] P --> B["Build a bundle roles first, then close matches, up to your size (10 by default)"] B --> F["Keep only what this agent can actually run"] F --> R["Routed set with exact files to read"]

• One index for every agent on your machine, deduplicated, and refreshed automatically when you install, remove or update a skill.
• Reads what each skill is about, not just its one-line description: the most distinctive words of every skill's body are indexed at rebuild.
• Bundles, not long lists. A route fills complementary roles (primary method, context, integration, verification, support), then tops up with close matches only, never past your bundle size (10 by default, 3 to 20).
• Optional upgrades, never required: an embedding sidecar for semantic re-ranking, and a decision-model stage (for example Laya or a cross-encoder) that starts in shadow mode so you can measure it before trusting it.

Prompt-injection firewall for skills and MCP

Skills and plugins are instructions your agent follows. Lockkeeper's scanner finds text that tries to override the agent, commands that send secrets or files to the network, credential-store access, code that decodes and runs hidden payloads, destructive commands, and invisible Unicode, across Markdown, configs and scripts.

• CI-ready verdicts: clean, suspect, hostile with exit codes 0, 1, 2.
• Live protection: a Claude Code hook blocks hostile tool calls be