Choose your MCP client
Add one server URL and Tool Search is on. Nothing to install, no tool list to maintain.
The Best
Tool Discovery
for AI Agents
Two tools instead of hundreds. Your agent searches by meaning, loads only the actions it needs, and keeps its context for reasoning.
#1 Tool search on ToolRet-full44,453 tools · 7,961 queries
93.8% Search accuracyMetaTool · nDCG@5
3.8× Cheaper per requestTool Search vs every tool loaded · 824 tasks
Give Maya the approved Finance access and update her IT request.
Two tools do the work
tool_search found 5 of 247 actions tool_execute ran all 5 on StackOne Done. Maya's group membership is updated in Okta. Her existing ServiceNow request records the change.
5 actions executed · 97% fewer tokens on tool definitions
Companies already connected through StackOne
Semantic tool search benchmarks. Custom-trained for
better tool selection.
Our fine-tuned retrieval model leads three public tool-retrieval benchmarks against Anthropic Tool Use, AgentCore Gateway and keyword search.
MetaTool
A public benchmark: 200 single-tool APIs and 2,000 queries, the smallest of the three, testing basic retrieval quality.
Benchmark on GitHub ↗ (opens in new tab)ToolBench
A public benchmark from OpenBMB: 10,439 real-world API tools and 451 queries across three instruction types.
Benchmark on GitHub ↗ (opens in new tab)ToolRet-full
A public tool-retrieval benchmark: 44,453 tools across code, web and custom sets, and 7,961 queries with graded relevance.
Benchmark on GitHub ↗ (opens in new tab) ·
| Benchmark | Metric | Method | Score | Source |
|---|---|---|---|---|
| MetaTool | nDCG@5 | StackOne | 93.8% | Measured, 2,000 queries |
| MetaTool | nDCG@5 | Anthropic Tool Use | 91.3% | Measured, 30-query sample |
| MetaTool | nDCG@5 | AgentCore Gateway | 72.8% | Measured, 2,000 queries |
| MetaTool | nDCG@5 | BM25 | 38.5% | Measured, 2,000 queries |
| ToolBench | Hit@5 | StackOne | 74.3% | Measured, 451 queries |
| ToolBench | Hit@5 | Anthropic Tool Use | 40.0% | Measured, 10-query sample |
| ToolBench | Hit@5 | AgentCore Gateway | 34.1% | Measured, 451 queries |
| ToolBench | Hit@5 | BM25 | 26.6% | Measured, 451 queries |
| ToolRet-full | nDCG@10 | StackOne | 54.4% | Measured, 7,961 queries |
| ToolRet-full | nDCG@10 | Qwen3-8B | 46.2% | Published by the benchmark authors |
| ToolRet-full | nDCG@10 | BM25s | 36.4% | Published by the benchmark authors |
| ToolRet-full | nDCG@10 | AgentCore Gateway | 22.8% | Measured, 7,961 queries |
MetaTool 93.8% nDCG@5 against Anthropic Tool Use 91.3%; ToolBench 74.3% Hit@5 against Anthropic Tool Use 40.0%; ToolRet-full 54.4% nDCG@10 against Qwen3-8B 46.2%
How MCP tool search works. From 1,000 agent tools
to 2
tool_search finds the right actions. tool_execute runs them. Your agent loads what the task needs, leaving more context for reasoning.
Your agentconnected to
Workday
Okta
ServiceNowYour agent's context
Without Tool Search
≈74ktokens for tools and results
With Tool Search
≈900tokens for tools and results
82×Send a prompt to watch the context fill.
What can I help with?
Pick an example below and send it.
Interactive simulation. Pick a prompt, then send it.
Zendesk ticket 1842 to Engineering and notify the incident channel.Check MMaya's finance onboarding and route any mismatch to HR.Verify SSam's hiring and HR records, then send People Ops the onboarding handoff.Check BILL-1042 against PO-2087 and route any mismatch to AP.Check AAcme's overdue invoice and create a follow-up for its account owner. You:
Give MMaya the approved Finance access and update her IT request.
I'll find the right actions, then work through the request.
Request context
An approved ServiceNow request identifies Workday employee 21001 and the Finance - NetSuite access group.
- Verify the employeeWorkday
tool_searchFinding the action…Found
workday_list_workerstool_executeMaya is an active employee; work email verified. - Match her identityOkta
tool_searchFinding the action…Found
okta_list_userstool_executeFound the Okta account matching her work email. - Find the approved groupOkta
tool_searchFinding the action…Found
okta_list_groupstool_executeFound the approved Finance group, managed in Okta. - Apply the approved accessOkta
tool_searchFinding the action…Found
okta_add_user_to_grouptool_executeAdded Maya to the approved group. - Record the changeServiceNow
tool_searchFinding the action…Found
servicenow_update_tasktool_executeRecorded the change in the existing IT request.
Done.
Maya's group membership is updated in Okta. Her existing ServiceNow request records the change.
Only 5 action schemas loaded, as they were needed.
You:
Escalate Zendesk ticket 1842 to Engineering and notify the incident channel.
I'll find the right actions, then work through the request.
Request context
The request supplies the Zendesk ticket ID, Jira project key and issue type, and Slack incident channel ID. The agent carries the ticket reference into the engineering issue and notification.
- Read the customer issueZendesk
tool_searchFinding the action…Found
zendesk_get_ticketstool_executeRetrieved the ticket description, priority and customer impact. - Create the engineering issue
Jiratool_searchFinding the action…Found
jira_create_issuetool_executeCreated a Jira issue with the customer impact and Zendesk reference. - Notify the incident channel
Slacktool_searchFinding the action…Found
slack_send_messagetool_executePosted the escalation with links to both records.
Done.
Engineering has a Jira issue linked to the Zendesk ticket, and the incident channel has the escalation details.
Only 3 action schemas loaded, as they were needed.
You:
Check MMaya's finance onboarding and route any mismatch to HR.
I'll find the right actions, then work through the request.
Request context
The onboarding request supplies Workday employee 21001 and NetSuite employee 846. The agent verifies they refer to the same person before comparing records.
- Read the HR recordWorkday
tool_searchFinding the action…Found
workday_list_workerstool_executeMaya's HR record lists Finance as her department. - Check the finance profile
NetSuitetool_searchFinding the action…Found
netsuite_get_employeetool_executeSame employee, but the department differs from Workday. - Resolve the HR teamServiceNow
tool_searchFinding the action…Found
servicenow_list_groupstool_executeFound the HR Operations assignment group. - Route the discrepancyServiceNow
tool_searchFinding the action…Found
servicenow_create_tasktool_executeReview task assigned to HR Operations
Done.
A department mismatch is recorded in a ServiceNow task assigned to HR Operations, with evidence from both employee records.
Only 4 action schemas loaded, as they were needed.
You:
Verify SSam's hiring and HR records, then send People Ops the onboarding handoff.
I'll find the right actions, then work through the request.
Request context
The approved handoff supplies a Greenhouse candidate ID, BambooHR employee ID and People Ops Slack channel ID. The agent verifies the records refer to the same hire before summarizing them.
- Find the approved candidate
Greenhousetool_searchFinding the action…Found
greenhouse_list_candidatestool_executeFound Sam using the candidate ID from the handoff. - Confirm the hiring statusGreenhouse
tool_searchFinding the action…Found
greenhouse_list_applicationstool_executeThe matching application is marked hired. - Check the HR record
BambooHRtool_searchFinding the action…Found
bamboohr_get_employeetool_executeMatched the approved hire to the employee record and start date. - Send the handoffSlack
tool_searchFinding the action…Found
slack_send_messagetool_executePosted the verified hiring and HR details to People Ops.
Done.
People Ops receives the confirmed hiring status and HR record needed to start Sam's onboarding.
Only 4 action schemas loaded, as they were needed.
You:
Check BILL-1042 against PO-2087 and route any mismatch to AP.
I'll find the right actions, then work through the request.
Request context
The agent resolves document numbers to NetSuite internal IDs, compares the bill and purchase order, then routes the exception to Accounts Payable.
- Resolve the documentsNetSuite
tool_searchFinding the action…Found
netsuite_execute_suiteql_querytool_executeFound the matching bill and purchase order in NetSuite. - Read the vendor billNetSuite
tool_searchFinding the action…Found
netsuite_get_vendor_billtool_executeRetrieved the vendor bill and its line items. - Compare the purchase orderNetSuite
tool_searchFinding the action…Found
netsuite_get_purchase_ordertool_executeSame vendor, but a line-item amount does not match. - Find Accounts PayableServiceNow
tool_searchFinding the action…Found
servicenow_list_groupstool_executeFound the Accounts Payable assignment group. - Route the exceptionServiceNow
tool_searchFinding the action…Found
servicenow_create_tasktool_executeException task assigned to Accounts Payable
Done.
A bill-line mismatch is recorded in a ServiceNow task for Accounts Payable, with references to the bill and purchase order.
Only 5 action schemas loaded, as they were needed.
You:
Check AAcme's overdue invoice and create a follow-up for its account owner.
I'll find the right actions, then work through the request.
Request context
The collection request supplies a Xero invoice ID, the matching Salesforce account ID and the Finance Slack channel ID. The agent checks the customer identity and due date before creating the follow-up.
- Check the invoice
Xerotool_searchFinding the action…Found
xero_get_invoicetool_executeThe invoice is overdue and still has an outstanding balance. - Find the account owner
Salesforcetool_searchFinding the action…Found
salesforce_get_accounttool_executeVerified the matching customer and read the account owner ID. - Create the follow-upSalesforce
tool_searchFinding the action…Found
salesforce_create_tasktool_executeAssigned a collection task to the account owner with the invoice reference. - Update FinanceSlack
tool_searchFinding the action…Found
slack_send_messagetool_executePosted the balance and collection-task reference to Finance.
Done.
The Salesforce account owner has a collection task, and Finance has the invoice balance and follow-up reference.
Only 4 action schemas loaded, as they were needed.
Context cost, measured. Tool Search keeps context flat
as your catalog grows.
We counted the input tokens of a tool-selection request across 47 StackOne connectors, 2,877 tools in total. Loading all of a connector's tools costs about 400 tokens per tool, on every turn. With tool_search, the agent reads a short list of matches instead, whatever the connector size.
Input tokens per request, by connector
- ≈400
- input tokens per tool with every tool loaded, averaged over 2,877 real tools, the short selection prompt included
- 1.1M+
- tokens to load all 47 connectors at once, more than a 1M-token context window holds
- 3.8×
- cheaper per request with Tool Search, at the same selection accuracy (92.6% vs 91.5%)
| Connector | Tools | Every tool loaded | Tool Search |
|---|---|---|---|
| Google Docs | 4 | 21,991 | 21,028 |
| Google Tasks | 12 | 3,135 | 3,522 |
| Browserbase | 17 | 2,731 | 2,636 |
| Calendly | 21 | 5,417 | 3,827 |
| CharlieHR | 22 | 4,086 | 2,845 |
| Datadog | 23 | 13,754 | 6,389 |
| Airtable | 24 | 13,053 | 6,175 |
| Databricks | 26 | 9,479 | 4,684 |
| EasyLlama | 26 | 7,331 | 4,104 |
| Zoom | 27 | 14,464 | 6,251 |
| Google Analytics | 29 | 26,598 | 7,777 |
| Lattice | 32 | 6,751 | 3,319 |
| Okta | 32 | 16,931 | 6,412 |
| Amplitude | 35 | 10,114 | 3,184 |
| Gong | 35 | 10,793 | 4,236 |
| Firebase | 36 | 32,263 | 10,979 |
| Google Calendar | 37 | 27,110 | 13,737 |
| Slack | 38 | 9,807 | 4,409 |
| Orderspace | 42 | 16,268 | 5,792 |
| Figma | 44 | 15,620 | 5,632 |
| Cezanne HR | 46 | 13,003 | 4,862 |
| Lever | 46 | 14,252 | 4,648 |
| Paylocity | 49 | 35,341 | 12,193 |
| Adobe Learning Manager | 50 | 16,341 | 4,593 |
| Todoist | 52 | 10,600 | 3,578 |
| Docebo | 53 | 25,515 | 5,880 |
| Personio | 54 | 27,477 | 6,982 |
| Attio | 58 | 20,568 | 5,652 |
| Microsoft Teams | 61 | 22,234 | 5,482 |
| GoCardless | 66 | 38,154 | 7,390 |
| Dynamics 365 Business Central | 70 | 19,730 | 5,143 |
| Gusto | 72 | 17,357 | 3,441 |
| Outlook Calendar | 77 | 29,781 | 4,376 |
| 360Learning | 78 | 28,000 | 4,147 |
| Mailchimp | 80 | 17,183 | 4,416 |
| HubSpot | 81 | 28,043 | 5,361 |
| Lokalise | 82 | 19,157 | 3,694 |
| Pinpoint | 85 | 57,214 | 9,954 |
| HiBob | 91 | 28,579 | 4,041 |
| QuickBooks Online | 95 | 36,516 | 8,829 |
| Humaans | 96 | 28,870 | 6,331 |
| Teamtailor | 100 | 26,764 | 4,769 |
| Workday | 114 | 109,447 | 5,548 |
| Ashby | 121 | 43,282 | 4,840 |
| SmartRecruiters | 153 | 47,691 | 7,040 |
| JobAdder | 179 | 50,968 | 5,215 |
| Vincere | 206 | 68,159 | 4,685 |
- Every tool loaded
- Tool Search
Accuracy with and without Tool Search. Same accuracy,
a fraction of the tokens.
On single connectors, agents using Tool Search pick the right action as often as agents given every tool. On larger tool libraries, published results show accuracy falling when every tool is loaded and recovering with tool search.
Right action, first try
Claude Sonnet 4.6 · StackOne benchmark
Tasks completed
Gemini 3.5 Flash · StackOne benchmark
Large tool libraries
Every wrong call is another model turn you pay for. Sources: Anthropic, Introducing advanced tool use, Nov 2025 (opens in new tab) · Gan and Sun, RAG-MCP, arXiv 2505.03275, May 2025 (opens in new tab)
- Every tool loaded
- Tool Search
Dynamic tool discovery with Weave. The right tools.
And the path between them.
Advanced Tool Search pairs fine-tuned embeddings with Weave, our action graph built from connector data and past executions. It maps which action supplies each input across apps, so the agent gets the path and not just the target, and finishes in fewer turns.
It runs the paths behind employee onboarding, ticket triage and deal risk scoring.
Input dependencyFrequently pairedTarget action
Deferred tool loading in any agent. Connect once and
leave Tool Search on.
Add StackOne to your MCP client. Tool Search is enabled by default.
MCP server URL
Paste this URL into your MCP client.
- Add the MCP server URL.
- Sign in and choose your accounts.
- Leave Load tools when needed on.
Using a token-based MCP URL?
Add tool-mode=search_execute to your existing api.stackone.com/mcp URL. Developer guide (opens in new tab).
Enabled by default
Select accounts
Choose the accounts and actions Claude can access.
Accounts Select All (3)
Search Accounts...
Link Account
Beam Notion • 35 Actions • Updated Jul 28, 2026 
Beam Salesforce • 380 Actions • Updated Jul 28, 2026 
Beam SAP S/4HANA • 117 Actions • Updated Jul 28, 2026 Load tools when needed
Your agent looks up the right action when it needs one, instead of loading
all 532 up front. Keeps it fast and inside its context limit.
Back
Authorize
3 accounts selected 532 actions selected
On by defaultTurned off
Load tools when needed Your agent looks up the right action when it needs one, instead of loading every action up front. Keeps it fast and inside its context limit. Every action loads into the agent up front. More of the context window goes on definitions, but the agent sees every tool at once.
Works in your MCP client
or your own agent.
Build with the StackOne SDK
The same two tools, tool_search and tool_execute, in your own agent loop with any framework.
import { StackOneToolSet } from "@stackone/ai";
const toolset = new StackOneToolSet({ search: {} });
// Two tools reach the model: tool_search and tool_execute
const tools = await toolset.openai({
mode: "search_and_execute",
accountIds: ["acc_123"],
}); Install
npm install @stackone/ai from stackone_ai import StackOneToolSet
toolset = StackOneToolSet(search={"method": "auto"})
# Two tools reach the model: tool_search and tool_execute
tools = toolset.openai(
mode="search_and_execute",
account_ids=["acc_123"],
) Install
pip install stackone-ai Guardrails for enterprise tool access. Your agent acts and
you keep control.
From discovery to execution, the connection to your systems stays managed by StackOne.
You set the scope
Access your team controls
Choose the connected accounts and enabled actions your agent can use. Search works within that configured access.
91.6% vs 57.3%Right action ranked first by the retriever, before the model chooses: scoped to one connector, against all 998 tools at once. StackOne retrieval benchmark, 1,843 tasks.
Explore Tool PoliciesStackOne handles the calls
Execution handled by StackOne
Your agent chooses the action and parameters. StackOne handles provider authentication and runs the API calls on the server.
Agent Execution EngineScreen the response
Protect the model's context
Configure Defender to screen returned tool content for prompt injection before it reaches the model.
Explore response protectionAI Tool Search Resources
Blog
Building Semantic Search for 34,000+ Actions
How enriched embeddings reach 84% Hit@5 across 9,340 actions, and why reranking did not help.
Blog
MCP Tool Search: BM25, TF-IDF, and Hybrid
The evaluation behind our local fallback search: 2,700 test cases across 270 tools.
Blog
How to Reduce MCP Token Usage by 96 to 99%
Four strategies, from schema compression to code execution, and the 460x reduction from search-first discovery.
Blog
Why AI Agents Kill Their Own Context Windows
How tool schemas, API responses, and multi-step workflows compound to exhaust your agent's reasoning space.
Frequently Asked Questions
What is a context window in AI, and why does it matter for tool use?
A context window is the token budget an LLM has for a request, including instructions, tool definitions, messages and results. Loading an entire tool catalog uses that budget before the agent starts reasoning. StackOne Tool Discovery loads action schemas as needed. See Advanced Tool Search in the docs. To shrink what tool results return, Deep Query filters and aggregates synced records before they reach the agent.
How does semantic tool search work?
Semantic tool search uses a fine-tuned bi-encoder model to match natural language queries against tool descriptions by meaning, not keywords. For example, the query "give an employee access to the Finance group" can match
okta_add_user_to_group. In our connector-scoped evaluation on connectors held out from training, with the connector known, the top result was the correct tool 92.8% of the time. This measures tool retrieval, not successful completion of the full task.Does tool search reduce accuracy?
Not in our benchmarks. Across 47 StackOne connectors, Claude Sonnet 4.6 picked the right action first 92.6% of the time with tool search and 91.5% with every tool loaded, while each request cost 3.8 times less. On 634 multi-step tasks, pass rates were 92.3% and 93.2% with 55% fewer input tokens. With large tool libraries, published results show tool search improving accuracy: Anthropic reports Claude Opus 4 accuracy on its MCP evaluations rising from 49% to 74% when tool search is enabled.
How is semantic search different from keyword search like BM25?
Keyword search like BM25 matches terms in the query and tool description. StackOne trains its semantic model to match the requested task even when the wording differs from the action name. The SDK's
auto mode uses semantic first and falls back to BM25+TF-IDF locally if the API is unreachable.What is StackOne Weave?
Weave is the dependency graph behind StackOne tool search. It maps input sources ahead of time so a result can include the actions needed to call it. For example,
okta_add_user_to_group needs an Okta userId and groupId; okta_list_users and okta_list_groups supply them. The agent still plans and executes the task, including matching identities across connectors.How do I enable tool search on my StackOne MCP server?
Advanced Tool Search is on by default. Paste
https://mcp.stackone.com/mcp into your client, sign in, and keep the Load tools when needed toggle on when you authorize. Your agent then sees two tools, tool_search and tool_execute, instead of hundreds of individual definitions. On token-based URLs, add tool-mode=search_execute to api.stackone.com/mcp instead. Neither route needs a migration or code changes. See StackOne MCP.Does StackOne MCP tool search work with Claude, ChatGPT, and Cursor?
Yes. Paste
https://mcp.stackone.com/mcp into any MCP client, including Claude, ChatGPT, Cursor, Windsurf, VS Code and Zed, then sign in with OAuth and authorize. Tool search is on by default, so the agent sees tool_search and tool_execute instead of hundreds of definitions. No SDK or code changes needed on the client side. See StackOne MCP for client guides.When should I NOT use tool search?
If your agent uses fewer than 10 tools, static definitions work fine: context cost is negligible and tool availability is deterministic. Tool search adds value when connecting multiple systems where tool count exceeds what fits in the context window, or when you need semantic search across connectors.
What are the current limitations?
StackOne semantic tool search requires network connectivity. The SDK can fall back to local keyword search, which may return different results. Search retrieves possible actions; the agent still supplies parameters, verifies records and handles action results. Code Mode is a separate execution capability.
Let your agent find the tools it needs.
Connect your enterprise apps and put Tool Search to work.