mikrotik-mcp

Context Length Optimization

MikroTik MCP ships 162 tools. At full verbosity the tool schema can occupy ≈ 54 000 tokens — more than the entire context window of many local LLMs (LM Studio, Ollama, etc.).

Starting with this release every tool carries a short title annotation (MCP spec 2025-03-26) and a trimmed description, reducing the description token budget by ~75 %.

---

What changed

1. title in ToolAnnotations

Every @mcp.tool() call now passes annotations=annotate(READ|WRITE|…, "Short Title").

# before
@mcp.tool(name="create_queue_type", annotations=WRITE)

# after
@mcp.tool(name="create_queue_type", annotations=annotate(WRITE, "Create Queue Type"))

The title field is part of the MCP Tool Annotations spec.
MCP clients that support it can display a compact tool list using only the title — without rendering the full description text — significantly shrinking the prompt context they pass to the LLM.

2. Trimmed descriptions

Verbose multi-paragraph docstrings with redundant Args: / Returns: sections have been replaced by concise one-liners. The function signature already carries full type information; the description only needs to convey what the tool does.

Tool	Before	After
create_queue_type	1 580 chars	145 chars
generate_wireguard_client_config	1 519 chars	91 chars
create_filter_rule	1 080 chars	≤ 120 chars
all 162 tools (avg)	209 chars	51 chars

Estimated description-token savings: ≈ 6 400 tokens (≈ 75 % reduction).

---

How to benefit

MCP clients that use title

Any client following the MCP spec can read tool.annotations.title and show that short string in its tool list rather than the full description. If you are building an integration, prefer displaying the title in compact views.

Local LLMs with small context windows

Even without client-side filtering, the trimmed descriptions directly reduce the tokens consumed at MCP initialisation. For a 64 k-token LLM this brings the tool schema from ≈ 54 k tokens → ≈ 48 k tokens, leaving meaningful room for the conversation.

Further reduction: If you only need a subset of tools (e.g. only DNS and WireGuard), you can comment out the unused scope imports in src/mcp_mikrotik/app.py to drop those tools entirely.

---

Developer notes

The annotate() helper lives in src/mcp_mikrotik/app.py:

def annotate(base: ToolAnnotations, title: str) -> ToolAnnotations:
    """Return a copy of *base* with a human-readable *title* attached."""
    return ToolAnnotations(
        title=title,
        readOnlyHint=base.readOnlyHint,
        destructiveHint=base.destructiveHint,
        idempotentHint=base.idempotentHint,
        openWorldHint=base.openWorldHint,
    )

When adding new tools, always use annotate() instead of a bare annotation constant:

# ✅ correct
@mcp.tool(name="my_new_tool", annotations=annotate(READ, "My New Tool"))

# ❌ avoid — omits the title
@mcp.tool(name="my_new_tool", annotations=READ)

This site is open source. Improve this page.