What is MCP tool bloat?
Tool bloat is the context an AI model spends on tool definitions before it reads your request. Every MCP tool arrives as a name, a description and a JSON schema, and many clients load all of them at the start of every conversation.
In Anthropic's example, 58 tools from five MCP servers took about 55,000 tokens before the conversation started 1. Accuracy suffers too: Claude's tool choice degrades once more than 30–50 tools are available 2.
How does on-demand tool search work?
Sidechat's MCP server answers tools/list with exactly two tools. search_tools takes a plain-language query and returns up to 20 ranked matches, each with its name, description and full input schema. execute_tool runs one of them by name.
The technique is called progressive disclosure, or dynamic tool discovery. The MCP project's own client guide recommends a search_tools meta-tool for large catalogs 3, and OpenAI and Cloudflare report large savings from the same idea 4,5.
→ search_tools({ "query": "publish website" })
← {
"tools": [
{
"name": "sites_publish",
"description": "…",
"inputSchema": { "type": "object", "properties": { … } }
},
… up to 20 ranked matches
]
}
→ execute_tool({ "tool": "sites_publish", "args": { … } })How many tokens does it save?
We counted Sidechat's tool definitions with the o200k_base tokenizer in October 2026. Listing the full catalog up front would take 189 times as many tokens as the two tools. Searches cost tokens too, so the saving is largest when a task needs a handful of tools out of hundreds.
| What your assistant loads | Tools | Tokens |
|---|---|---|
| Sidechat's tools/list | 2 | 648 |
| Server instructions, sent once per session | – | 451 |
| One search with 20 results, typical | 20 | 4,600 to 6,300 |
| Every tool on a standard workspace, listed up front | 465 | 122,572 |
When is tool search the wrong choice?
If your agent only ever uses a few tools, list them directly. Anthropic recommends tool search from about 10 tools or 10,000 tokens of definitions 2; below that, the extra search call buys little.
Search also adds a step. The first use of a tool takes a search call before the tool call, which adds a round trip in exchange for a smaller context on every turn.
What about work that repeats every day?
Agents are expensive to run on a loop: they typically use about 4 times the tokens of a chat, and multi-agent systems about 15 times 6.
Once a workflow is settled, your assistant can turn it into a scheduled Function on Sidechat: code that runs on a schedule and calls a model only where the code asks for one. A daily report that fetches, filters and posts results makes no model call unless one of its steps needs judgment.
How do you connect your AI assistant?
Once your workspace has access, connecting takes one command. Your assistant discovers the rest of the catalog as it works.
- 1
Add the server
In Claude Code, run
claude mcp add --transport http sidechat https://canopy.side.chat/mcp. In claude.ai, Codex, Cursor or VS Code Copilot, add the same URL as a custom MCP server. - 2
Sign in
Your client opens a Sidechat sign-in page. Tool calls then run as you, so arguments never carry organization or user IDs.
- 3
Ask for the job, not the tool
Your assistant calls
search_toolswith plain words, such as “publish website”, and gets ranked matches with their input schemas. - 4
Call the tool it picked
execute_toolruns the chosen tool by name with its arguments.
MCP tool search questions
Does it work with my MCP client?
Sidechat's MCP server is documented for Claude Code, claude.ai custom connectors, Codex, Cursor and VS Code Copilot. Any MCP client that supports remote servers with OAuth sign-in should work; Claude Code is verified end to end.
Does searching add a step?
Yes. The first time your assistant needs a tool, it makes one search call before the tool call. In return, it doesn't carry hundreds of definitions in every turn.
How many tokens does a search cost?
A typical search that returns 20 matches costs about 4,600 to 6,300 tokens, because each match includes its full input schema.
How many tools are in the catalog?
A standard workspace has 465 tools across the default plugins. The number grows as more plugins are enabled.
Do tool calls need organization or user IDs?
No. Identity comes from your sign-in, so arguments never carry organization or user IDs.
How did you count the tokens?
Each tool's name, description and input schema was serialized the way an MCP client receives it and counted with the o200k_base tokenizer in October 2026. Claude's current tokenizer isn't public, so treat the absolute numbers as about ±10%; the ratio holds.
Keep reading
Sources
- In Anthropic's example, 58 tools from five common MCP servers took about 55,000 tokens before a conversation started. Tool search cut a 77,000-token setup to 8,700 tokens and improved Claude Opus 4.5's accuracy from 79.5% to 88.1%. Anthropic, Introducing advanced tool use, .
- Claude's ability to pick the right tool degrades once more than 30–50 tools are available. Anthropic recommends tool search from about 10 tools, or 10,000 tokens of definitions. Claude docs, Tool search tool, .
- The MCP project's client guide recommends a lightweight search_tools meta-tool, and says loading every tool definition up front wastes tokens, adds latency and degrades model performance. Model Context Protocol, Client best practices, .
- OpenAI reported that tool search cut total token use by 47% at the same accuracy, across 250 tasks on 36 MCP servers. VentureBeat, on OpenAI's GPT-5.4 launch, .
- Cloudflare put more than 2,500 API endpoints behind two tools, search() and execute(), using about 1,000 tokens where an equivalent MCP server would need 1.17 million. Cloudflare, Code Mode: give agents an entire API in 1,000 tokens, .
- Agents typically use about 4 times the tokens of a chat interaction, and multi-agent systems about 15 times. Anthropic, How we built our multi-agent research system, .
Changes to this page
- : First published.
- : Added measured token counts for Sidechat's catalog, how tool search works, its tradeoffs, sources and questions.