Today, Arcade.dev is releasing Tool Recommendation, a new capability within Arcade that gives agents your whole tool catalog at a fraction of the token cost.
We built Tool Recommendation so you can save money and tokens, stop maintaining a tool list for every agent, and have better odds of your agent finding the right tool. Specifically, that means up to 91% fewer input tokens per session and up to 37% lower token costs, with industry-leading odds of your agent finding the right tool. These results hold across any model.
Tool Recommendation matches each task to the few tools your agent needs, at the moment it needs them, so your agent gets the right tool without carrying thousands of definitions it will never call.
That efficiency matters most in complex work where a single request can chain a dozen tool calls. Since every turn now runs on a lighter context, you stop paying for tools your agent never uses, removing the need to limit tool use upfront and unlocking the multi-step, multi-system workflows that businesses depend on.
Critically, Tool Recommendation is built into the Arcade actions runtime so every tool call benefits from our trusted governance and policy enforcement. Tool Recommendation alone saves tokens. Tool Recommendation inside an actions runtime makes the whole catalog safe to use.
The problem: agents hit a tool wall at enterprise scale
An agent earns its keep by acting in the systems people use every day: the CRM, the ticket queue, the ERP, the code repo, the inbox. However, those systems carry more tools than most teams expect, because every action an agent can take is its own tool. In Arcade’s tool catalog, Gmail alone has 30: drafting an email is one tool, sending it is another, forwarding a third, with more for searching, labeling, replying, and triage. Similarly, Salesforce carries 17 tools. GitHub carries 43 and Jira carries 45. Connect every system an enterprise runs and the tool count climbs into the thousands.
Using Arcade’s 8,000+ tool catalog as an example, loading the catalog would cost over 1.75M tokens before the agent reads a single message. That catalog size is also more than the context window (the amount of text a model can process in a single request) of any major model, so even if you were willing to eat the cost, loading entire catalogs wouldn’t work. And accuracy drops well before the window fills, because a model choosing from thousands of options picks the wrong tool more often.
Currently, teams cope in one of three ways.
- Curate a short tool list per agent, which caps what the agent and the end-user can do and turns every new tool into a ticket.
- Ship the whole catalog and accept the wrong tool calls, until the list outgrows the model’s context window and stops working.
- Swap in a different tool list for each task, which means someone has to build and maintain the logic that decides which list fits while managing an ever-increasing list of new tools.
Each route gives up capability, cost, or efficiency, and Gartner predicts that inference costs per agentic workflow will rise more than fivefold through 2028.
We hit this tool wall in our own work at Arcade while simultaneously watching our enterprise customers run into it at the same time. So we set out to fix the problem while holding to three principles:
- The number of tools should never be a limiting factor for agents.
- There needs to be one solution that works across models, agents, clients, etc.
- Every action needs to be fully governed.
Tool Recommendation is our answer. After more than a year of work, countless iterations, optimizations, and tests against industry-standard benchmarks, we’re now proud to make it available.
The solution: how Tool Recommendation works
Tool Recommendation turns tool selection into a retrieval problem instead of a reasoning problem, an idea with e-commerce roots. When you open Amazon, a recommendation engine powers the “you might also like” row. It pulls the best matches from a catalog of millions of products in the time the page takes to render, and it improves with every release. We thought, what if we applied the same approach to agent tools?
So instead of asking a model to reason across every tool, we built Tool Recommendation to retrieve the few that fit.
The agent asks for tools when they need them, instead of carrying them
The agent connects to an Arcade gateway that exposes a few built-in tools, including Arcade_ListApps, Arcade_SelectTools, and Arcade_UseTool, in place of the full catalog.
When a task arrives, the agent describes what it needs in plain language and calls Arcade_SelectTools. Arcade returns the few tools that fit, with their schemas, as a tool result. The agent then runs its pick through Arcade_UseTool.
A single request can chain a dozen of these steps: pull the account from Salesforce, check tickets in Zendesk, update the forecast in NetSuite, draft the renewal in Google Docs. Each step fetches only what it needs, so every turn carries a lighter context.
A smaller model does the fetching
Arcade does not ask a frontier model to pick from a list. Instead, purpose-built embedding and ranking models do the work. The embedding model matches the agent’s plain-language request against the name and description of every tool in the catalog, and the ranking model reorders the closest candidates so the best fits come first. These models only retrieve and generate nothing, so each query costs far less than an LLM call, and the agent’s model chooses from a short list that is already relevant instead of reasoning across thousands of definitions.
Tool descriptions drive recommendation quality
Because matching runs on names and descriptions, clear ones produce better matches and thin ones produce worse matches. Description quality is the one lever you control and should be an evaluation criteria when you’re selecting which MCP tools to use. For an objective benchmark of tool quality, see Arcade’s ToolBench.
Works with any model, framework, or MCP client
Tool Recommendation runs over MCP, so it works with any MCP client, in any framework, without SDK changes or custom integration. Whichever assistant or agent your teams already use, off-the-shelf or built in-house, gets the whole catalog the same way. The savings hold across every model we tested, and you turn it on per Arcade gateway, with nothing in your agents to rebuild.
The results: lower costs, fewer tokens, increased accuracy
The savings: up to 37% lower token costs
Return on investment has become the question every enterprise agent program has to answer. Token spend grows with the number of users, conversations, turns, and tools, so reducing token costs was our primary driver.
Compared with an assistant given a fixed list of 100 tools up front, Tool Recommendation saves about 20% of token costs in a typical conversation, and up to 37% depending on the model. At 1,000 seats on Claude Opus 5, that adds up to about $119K a year. The larger the catalog, the bigger the difference, so teams with thousands of tools stand to gain the most.
Up to 91% fewer tokens per session
Every token spent on tool definitions is context the model cannot spend on the task, and those definitions ride along on every turn. A leaner prefix leaves more room for the conversation and the results that matter.
Across a 10-turn session, models carry up to 91% fewer input tokens, on every model tested. A GPT-5.6 session drops from 256,650 tokens to 22,164. On Claude Opus 5, it drops from 411,729 to 38,170.
Better odds of finding the right tool
ROI extends well past cost. Enterprises need work done reliably, and an agent that picks the wrong tool wastes tokens, stalls the workflow, and loses the team’s trust.
Microsoft recently published results for Foundry’s tool search, giving the industry a reputable benchmark to measure against. We ran Arcade on ToolRet, the public academic benchmark for tool retrieval, and compared the results to Microsoft’s published numbers. Arcade surfaced the needed tools more often in three primary categories, measured by Recall@10.
For additional details on these results, see Notes on the Results.
Tool Recommendation as part of the Arcade actions runtime
Tool Recommendation runs inside the Arcade actions runtime, the layer between your agents and your business systems that authorizes, enforces, and governs every action an agent takes.
Authorization: every action runs as the user
Arcade authorizes each action at the intersection of what the user can do and what the agent is allowed to do, and checks both when the action happens. Users grant access through each provider’s own OAuth flow, the runtime requests only the scopes an action needs, and credentials stay in the Arcade vault, outside the model’s context.
Enforcement: your policies apply before and after every call
Arcade enforces the policies that already run your business. It integrates with the policy engines and identity providers your security team already uses, so teams set policy once and Arcade applies it through pre- and post-tool-call hooks on every action. Those hooks can also screen each call for personally identifiable information (PII) and apply data loss prevention (DLP) protection, so sensitive data is caught before it reaches the model. The access filter runs on every recommendation and every call, so no agent reaches a tool it is not cleared to run, however large the catalog grows. Security sets the rule once and can trust that it is enforced.
Governance: one audit trail for every tool, agent, and action
A central control plane answers whether a given agent, acting for a given user, can perform a given action on a given resource. Every recommendation and call is audited, and Arcade standardizes the logging regardless of agent, model, MCP tool, or client. The records feed the observability stack a team already runs, through OpenTelemetry, so security reads one consistent trail instead of a different log for every agent.
Deployment: run it in your environment or try it in Arcade Cloud
Tool Recommendation is available now in Arcade Cloud, so you can try it today. Many of our enterprise customers cannot send agent traffic outside their own environment, so it also runs in your environment, whether in your cloud, on-premises and in air-gapped environments. Wherever it runs, it connects to the identity providers and policy engines you already use, so it is not a silo your security team has to manage separately.
Get more from your agents for less
Agents should never be limited by the tools they can reach. With Tool Recommendation you get the whole tool catalog, so nobody curates a list per agent or files a ticket to add a tool, and the multi-step workflows businesses depend on become possible.
Because each agent asks only for the few tools a task needs, Tool Recommendation carries up to 91% fewer tokens per session and costs up to 37% less to run while finding the right tool with industry-leading accuracy. And you can use it with any model, framework, or client while fully governing every agent action through the Arcade actions runtime.
Tool Recommendation is available now, for all Arcade Cloud users, and we are rolling it out to enterprise customers. If you’d like to try Tool Recommendation in your environment, reach out to the team and we can get it set up.
More on Tool Recommendation
Notes on the Results
Savings
Compares an AI assistant using Tool Recommendation with the same assistant given a fixed list of 100 tools up front. Percentages and dollars are modeled for input cost only; output tokens are excluded. A typical conversation is 3 requests with about 8 tool calls each, with 0.12 searches per tool call (observed on Arcade’s plugin gateway) and a 100-tool list at the catalog’s average definition size. The model assumes the client keeps large tool results out of the conversation; if results stay inline, the saving is about 5% to 7%. Savings reach up to 37% in single-request sessions and shrink as conversations lengthen. A fixed list is cheaper below roughly 25 to 28 tools. Dollars use each vendor’s published prices as of 10/6/26. The $119K figure assumes 20 conversations per user per month (240,000 a year). In paired end-to-end runs against a 100-tool list (408 sessions across 64 workflows, using Codex and Claude Code), the total bill, input and output, fell 7.6% (95% interval 2.5% to 12.6%). Every one of the 14 runs was cheaper, and task completion did not change. The decline was 5.5% with Codex (interval includes zero) and 9.7% with Claude Code.
Token Use
Modeled for a 10-turn session, input tokens only, covering tool definitions and recommendation responses (excludes conversation and tool results). The baseline is the first 100 tools in the Arcade catalog, which average about 257 tokens per definition, about 20% more than the catalog average. A 100-tool list is the comparison because the full catalog (about 1.76M tokens) fits in no model’s context window. Counts come from 48 hours of production traffic (85 tool calls). OpenAI counts are exact, and Claude counts are estimated from measured calibration. These are definition-token savings, not whole-session savings. In paired end-to-end runs, total input tokens fell 17% with Codex and 36% with Claude Code.
Accuracy
Recall@10 is the share of needed tools that appear in the top 10 results. Lead is relative: Arcade’s score divided by Microsoft’s, minus one. Arcade’s figures come from our own run of ToolRet (7,961 queries over 44,453 tools) at Tool Recommendation’s production ranking settings, which include the cross-encoder reranker. Our run searches the full tool index using the instruction Tool Recommendation sends in production. Microsoft’s figures are quoted from its published results, which use ToolRet’s setting without instructions. Because the two sets of scores come from different test setups, the comparison is indicative rather than exact.
