Metadata-Version: 2.4
Name: clear-your-tools
Version: 0.0.9
Summary: Clear Your Tools (CYT) — dynamic tool gating for eliminating the MCP/tools tax
Project-URL: Homepage, https://github.com/qdrddr/clear-your-tools
Project-URL: Repository, https://github.com/qdrddr/clear-your-tools
Project-URL: Issues, https://github.com/qdrddr/clear-your-tools/issues
Author: Damien Berezenko
License-File: LICENSE
Keywords: CYT,Clear Your Tools,mcp,mcp-client,tool-gating
Classifier: Intended Audience :: Developers
Classifier: Operating System :: MacOS :: MacOS X
Classifier: Operating System :: Microsoft :: Windows
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.13
Requires-Python: <4.0,>=3.13
Requires-Dist: pydantic==2.12.5
Requires-Dist: python-dotenv==1.2.2
Requires-Dist: pyyaml>=6.0.2
Requires-Dist: tiktoken==0.12.0
Requires-Dist: typing-extensions==4.15.0
Provides-Extra: all
Requires-Dist: h2==4.3.0; extra == 'all'
Requires-Dist: httpx>=0.28.0; extra == 'all'
Requires-Dist: hypercorn==0.18.0; extra == 'all'
Requires-Dist: libsql-experimental>=0.0.55; extra == 'all'
Requires-Dist: litellm==1.83.14; extra == 'all'
Requires-Dist: starlette>=0.46.0; extra == 'all'
Requires-Dist: uvicorn>=0.34.0; extra == 'all'
Provides-Extra: proxy
Requires-Dist: h2==4.3.0; extra == 'proxy'
Requires-Dist: httpx>=0.28.0; extra == 'proxy'
Requires-Dist: hypercorn==0.18.0; extra == 'proxy'
Requires-Dist: libsql-experimental>=0.0.55; extra == 'proxy'
Requires-Dist: starlette>=0.46.0; extra == 'proxy'
Requires-Dist: uvicorn>=0.34.0; extra == 'proxy'
Provides-Extra: pruners
Requires-Dist: litellm==1.83.14; extra == 'pruners'
Description-Content-Type: text/markdown

# Clear Your Tools

**Clear Your Tools** is a reverse proxy for coding agents such as
[Claude Code](https://docs.anthropic.com/en/docs/claude-code). It sits between the agent and upstream
LLM providers (Anthropic-compatible APIs on OpenRouter, Novita, DeepInfra, and others), intercepts
each request, and shrinks the tool payload before forwarding it upstream. Can be easily adopted for
other harness agents.

Large MCP catalogs can add tens of thousands of tokens of tool-schema overhead on every turn.
Clear Your Tools removes irrelevant tools and trims irrelevant optional parameters while always
keeping required fields for tools that stay in the request.

---

## How it works

```text
Agent (Claude Code, etc.)
        │
        ▼
Clear Your Tools proxy  ──► extract user query from messages
        │                   decompose each tool schema
        │                   score / filter with reranker (or LLM pruning)
        │                   recompose pruned tool list
        ▼
Upstream provider (OpenRouter, Anthropic, Novita, …)
```

On each intercepted request the proxy:

1. **Extracts the user query** from the conversation (latest user turn, with message cleanup).
2. **Decomposes tool schemas** into a catalog of chunks: each tool root keeps required properties;
   optional properties are split into separate searchable units.
3. **Runs the pruning pipeline** configured in `config.yaml` (default: `rerank`; or `llm`).
4. **Recomposes surviving tools** — required properties always remain; only optional properties
   that look relevant to the query are merged back in.
5. **Forwards the modified request** to the upstream provider with the smaller `tools` array.

### Pruning pipeline

| Stage    | Model (default)                        | When it runs                                                          | What it does                                                                                     |
| -------- | -------------------------------------- | --------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------ |
| `rerank` | Qwen3-Reranker-8B (DeepInfra)          | ≥ `models.rerankers.minimum_tools` tools (default **29**)             | Scores every catalog chunk against the user query; drops low-scoring tools and optional props.   |
| `llm`    | Mercury 2 or GPT-OSS-120B (OpenRouter) | ≥ `models.llm.minimum_tools` tools (default **50**), after `rerank`   | LLM selects which catalog chunks to keep; can remove entire tools more aggressively.             |

**Recommendations:**

- **50+ tools** — keep **`rerank`** or use **`llm`**. rerank can be pipelined into LLM as a second
  stage (`pipeline: [rerank, llm]`) for stronger tool-level filtering on large catalogs.

---

## Quick start

Requires uv tool.
Install [uv](https://docs.astral.sh/uv/getting-started/installation)

### 1. Install proxy

From PyPI (proxy + pruners):

```bash
uv tool install 'clear-your-tools[all]'
```

<details>
<summary><strong>Though we strongly recommend using password vaults like macOS KeyChain</strong></summary>

```shell
# Store key in secure vault
security add-generic-password -s "nono" -a "OPENROUTER_API_KEY" -w "sk-..."  # macOS

# Now you can access the key like this:
export ANTHROPIC_AUTH_TOKEN="$(security find-generic-password -s "nono" -a "OPENROUTER_API_KEY" -w)"
```

</details>

### 2. Configure the proxy

Interactive wizard (writes `~/.config/cyt/config.yaml` and optionally `~/.config/cyt/.env`):

```bash
uv run cyt-rproxy setup
```

Or edit `~/.config/cyt/config.yaml` manually — see [CONFIG.md](CONFIG.md).

### 3. Run the proxy

Installed CLI:

```bash
uv run cyt-rproxy serve
```

Default listen port: **8834** (from bundled `defaults.yaml` or `~/.config/cyt/config.yaml`).

### 4. Run the the Agent

Point Claude Code at the proxy:

```bash
export ANTHROPIC_BASE_URL="http://localhost:8834/anthropic"
export OPENROUTER_API_KEY="..."
export ANTHROPIC_AUTH_TOKEN="${OPENROUTER_API_KEY}"
claude --model haiku 'say hi' -p
```

The default upstream in `config.yaml` is OpenRouter's Anthropic-compatible endpoint. Change
`network.proxy.reverse.upstreams` to target a different provider URL.

### 5. View pruning stats savings

```bash
uv run cyt-rproxy stats totals
uv run cyt-rproxy stats summary --period day
uv run cyt-rproxy stats events --limit 20
```

Stats are stored in `~/.config/cyt/stats.db` by default.

---

## FAQ

<details>
<summary><strong>Doesn't pruning burn more tokens than it saves?</strong></summary>

The reranker and weak LLM used for pruning are **much cheaper per token** than the main model
(e.g. Claude Sonnet). You may spend extra tokens on pruning, but they cost a fraction of what you
save on the main request. Set `input_cost_per_token` and `output_cost_per_token` in
[`~/.config/cyt/config.yaml`](CONFIG.md#configuration) to track savings.

**Example pricing (input tokens):**

| Model               | Cost per 1M input tokens |
| ------------------- | ------------------------ |
| Claude Sonnet 4.6   | $3.00                    |
| Qwen-Reranker-8B    | $0.050                   |
| GPT-OSS-120B        | $0.14                    |
| Inception Mercury 2 | $0.25                    |

The weak models such as Mercury 2 or GPT-OSS-120B returns only the IDs of tools to keep, so its
output stays extremely small. Rerankers do not count output tokens and are usually much cheaper
than a strong LLM.

**Rule of thumb:** saving 1M Sonnet input tokens is still worthwhile even if pruning uses up to
~10M Mercury tokens — roughly a 1:10 cost ratio. The reranker has roughly a 1:60 cost ratio.

In practice, pruning usually adds modest overhead. Worst case (no tools pruned), you might pay
~$3.30 instead of $3.00. With typical pruning (40–95% of tool tokens removed), tool-schema cost
drops from ~$3.00 to roughly **$0.15–$1.80**, plus ~$0.30 for pruning — about **$0.45–$2.10 total**
for tool-related cost, or roughly **30–85% savings** depending on policy.

</details>

<details>
<summary><strong>Why don't I see 30–85% savings on my total request?</strong></summary>

Those numbers apply to **tool schemas only** of the **input tokens only**, not the full prompt (system message, conversation
history, user message, etc.). Clear Your Tools prunes tools based on the user request; the rest of
the request is unchanged.

How much you save overall depends on:

- **How many tools you have** — more MCP servers mean a larger share of the request is tool
  schemas. We do not recommend using CYT below 50 tools.
- **Which pruning policy you use** — see [Pruning policies](CONFIG.md#configuration).

To estimate savings on a captured request JSON, see [`DEV.md`](DEV.md).
To see statistics of actual net savings (input tokens) run:

```bash
uv run cyt-rproxy stats totals
```

With ~100 tools and `prune_all`, expect **~85–95% savings on tool tokens** and typically **~30%+
savings on the full request**. The more tools you have the more overall savings you'll see.

</details>

<details>
<summary><strong>Where can I see how many tools and parameters an MCP server has?</strong></summary>

The popular [Fetch](https://mcpmarket.com/server/fetch) MCP server is a good example. On its
**Tools** tab: 4 tools, each with 4 parameters (1 required, 3 optional) — 16 parameters total.

If the user asks to "fetch the Markdown of a webpage", the `prune_all` typically keeps only the
**Fetch Markdown** tool with its required parameter plus any optional parameters that look
relevant. Unrelated tools (e.g. **Read file**) are dropped entirely.

</details>

---

## Development

See [`DEV.md`](DEV.md) for checkout setup, repository layout, library usage, and configuration reference.

---

## Limitations

See [`LIMITATIONS.md`](LIMITATIONS.md) for deployment constraints, token accounting caveats, and MCP aggregator trade-offs.

## Debug

See details to debug pruning in [debug/](debug/).

---

## License

<details>
<summary><strong>Inspiration</strong></summary>

This project is inspired by the ideas explored in the [tool-attention](https://github.com/asadani/tool-attention) project,
particularly around improving tool selection efficiency and reducing unnecessary tool exposure to the model.

It also aims to limit the effects of [context rot](https://www.trychroma.com/research/context-rot)
by pruning irrelevant or confusing tools from the available toolset based on the current user prompt and execution context.

Reducing irrelevant tools helps decrease prompt noise, lowers cognitive load on the model,
and can improve tool selection accuracy and overall agent reliability.

</details>

See [`LICENSE`](LICENSE).
