Skip to content

Commit d39d93b

Browse files
committed
Sync open source content 🐝 (from c6a5723e8e7e0e0fae964cba8912e29b5180cfdb)
1 parent 480db4b commit d39d93b

File tree

1 file changed

+39
-20
lines changed

1 file changed

+39
-20
lines changed

docs/gram/build-mcp/dynamic-toolsets.mdx

Lines changed: 39 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -5,54 +5,73 @@ description: Enable very large MCP servers by making a toolset dynamic
55

66
import { Callout } from "@/mdx/components";
77

8-
Dynamic toolsets enable very large MCP servers without overloading context windows. Instead of exposing all tools upfront like traditional MCP, dynamic toolsets provide "meta" tools that allow the LLM to discover only the tools it needs to complete specific tasks, optimizing token and context management.
8+
Dynamic toolsets enable very large MCP servers without overloading context windows. Instead of exposing all tools upfront like traditional MCP, dynamic toolsets provide "meta" tools that allow the LLM to discover only the tools it needs to complete specific tasks, delivering up to 160x token reduction while maintaining full functionality.
99

10-
Gram exposes two types of dynamic toolsets, both of which are experimental:
10+
Our refined Dynamic Toolsets approach combines the best of semantic search and progressive discovery into a unified system that exposes three core tools. For detailed technical insights and performance benchmarks, see our [blog post on how we reduced token usage by 100x](/blog/how-we-reduced-token-usage-by-100x-dynamic-toolsets-v2).
1111

12-
## Progressive Search
12+
## How Dynamic Toolsets work
1313

14-
Progressive Search uses a "progressive discovery" approach to surface tools. Tools are organized into groups that the LLM can inspect to gradually discover what tools are available to it. Details of tools are only exposed when needed, for example tool schemas (which represent a large portion of tool token use) are only surfaced when the LLM decides it actually wants to use a specific tool. The toolset is compressed into three tools that actually get exposed directly to the LLM:
14+
Dynamic toolsets follow the natural workflow an LLM needs: search → describe → execute. The system compresses large toolsets into three meta-tools:
1515

16-
### `list_tools`
16+
### `search_tools`
1717

18-
The LLM can discover available tools using prefix-based lookup (e.g., `list_tools(/hubspot/deals/*)`). This process is accelerated by providing the structure of available tools in the tool description, creating a hierarchy of available sources and tags. This allows the LLM full control over what tools it discovers and when.
18+
The LLM searches for relevant tools using natural language queries with embeddings-based semantic search. The tool description includes categorical overviews of available tools (e.g., "This toolset includes HubSpot CRM operations, deal management...") and supports filtering by tags like `source:hubspot` for precise discovery.
1919

2020
### `describe_tools`
2121

22-
The LLM can look up detailed information about specific tools, including input schemas. While this could be combined with `list_tools`, the input schemas represent a significant portion of tokens, so keeping them separate optimizes token and context management at the cost of speed.
22+
The LLM requests detailed schemas and documentation only for tools it intends to use. This separation optimizes token usage since input schemas represent 60-80% of total tokens in static toolsets.
2323

2424
### `execute_tool`
2525

26-
Execute the discovered and described tools as needed for the specific task.
26+
The LLM executes discovered and described tools with proper parameters.
2727

28-
## Semantic Search
28+
## Performance benefits
2929

30-
Semantic Search provides an embeddings-based approach to tool discovery. Embeddings are created in advance for all the tools in a toolset, then searched over to find relevant tools for a given task.
30+
Dynamic toolsets deliver significant advantages over static toolsets:
3131

32-
### `find_tools`
32+
**Massive token reduction**: Input tokens are reduced by an average of 96% for simple tasks and 91% for complex tasks, with total token usage dropping by 96% and 90% respectively.
3333

34-
The LLM can execute semantic search over embeddings created from all tools in the toolset, allowing for more intuitive tool discovery based on natural language descriptions of what it wants to accomplish. This is generally faster than Progressive Search especially for large toolsets, but has less complete coverage and may result in worse discovery. The LLM has no insight into what tools are available broadly and can only operate off of whatever the semantic search returns.
34+
**Consistent scaling**: Token usage remains relatively constant regardless of toolset size. A 400-tool dynamic toolset uses only ~8,000 tokens initially compared to 410,000+ for the same static toolset.
3535

36-
### `execute_tool`
36+
**Context window compatibility**: Large toolsets that exceed Claude's 200k context window limit with static approaches work seamlessly with dynamic toolsets.
37+
38+
**Perfect reliability**: Maintains 100% success rates across all toolset sizes and task complexities.
39+
40+
### Sample performance data
41+
42+
| Toolset Size | Mode | Simple Task Tokens | Tool Calls | Complex Task Tokens | Tool Calls |
43+
|-------------|------|-------------------|------------|-------------------|------------|
44+
| 100 tools | Static | 159,218 | 1 | 159,216 | 3 |
45+
| 100 tools | Dynamic | 8,401 | 3 | 18,095 | 7 |
46+
| 400 tools | Static | 410,738 | 1 | 410,661 | 3 |
47+
| 400 tools | Dynamic | 8,421 | 3 | 31,355 | 7.8 |
48+
49+
## Trade-offs
3750

38-
Execute the tools found through semantic search.
51+
While dynamic toolsets offer significant benefits, there are some considerations:
3952

40-
## Benefits
53+
**Increased tool calls**: Dynamic toolsets require 2-3x more tool calls (typically 6-8 for complex tasks vs 3 for static), following the search → describe → execute pattern.
4154

42-
Both dynamic toolset approaches share the same core benefit: they avoid dumping all tools into context upfront. Instead, they expose the LLM to only the tools actually needed for a given task, making it possible to work with very large toolsets while maintaining efficient context usage.
55+
**Potential latency**: Additional tool calls may introduce slight latency, though this is often offset by reduced token processing time.
4356

44-
This approach is particularly valuable when working with extensive APIs or large collections of tools where loading everything at once would exceed context limits or create unnecessary complexity.
57+
**Complexity**: The multi-step discovery process adds complexity compared to direct tool access, though this is handled automatically by the LLM.
4558

4659
## Enabling dynamic toolsets
4760

48-
Head to the `MCP` tab to switch your toolset to one of the above dynamic modes.
61+
Head to the **MCP** tab in your Gram dashboard and switch your toolset from "Static" to "Dynamic" mode.
4962

5063
<Callout title="Note" type="info">
51-
This setting only applies to MCP, and will not affect how your toolset is used in the playground.
64+
This setting only applies to MCP and will not affect how your toolset is used in the playground, where static tool exposure remains useful for testing and development.
5265
</Callout>
5366

54-
![enabling dynamic toolsets](/assets/docs/gram/img/dashboard/tool-selection-mode.png)
67+
Dynamic toolsets are particularly valuable for:
68+
- APIs with 100+ operations
69+
- Enterprise systems with comprehensive toolsets
70+
- Applications where context window limits are a concern
71+
- Production environments requiring predictable costs
5572

5673
## Additional reading
5774

75+
- [How we reduced token usage by 100x with Dynamic Toolsets](/blog/how-we-reduced-token-usage-by-100x-dynamic-toolsets-v2)
5876
- [Code Execution with MCP](https://www.anthropic.com/engineering/code-execution-with-mcp)
77+
- [Previous Dynamic Toolsets implementation](/blog/100x-token-reduction-dynamic-toolsets)

0 commit comments

Comments
 (0)