New Tool Recommendation system cuts your firm's input token costs by 37%, according to Arcade.dev

Arcade.dev, a startup that helps AI agents work inside business software while controlling what they are allowed to do, is introducing Tool Recommendation, a feature that helps agents find the right actions to take without repeatedly loading a lengthy list of software functions into their conversations.

The release targets a practical problem for enterprise AI: Connecting an agent to more business systems also makes its instructions larger and more expensive.

Arcade.dev, also known as "Arcade," says its approach reduces the tokens devoted to tool discovery by as much as 91%, with modeled input-cost savings reaching 37%.

In a separate test of complete workflows, the company reports a smaller but more directly applicable result: a 7.6% reduction in the combined input/output token bill.

The feature becomes available to customers of the firm's Arcade Cloud fully managed software-as-a-service platform on October 15 without a separate feature charge, an Arcade spokesperson told VentureBeat.

The company's normal platform and usage fees still apply to paid plans, while enterprise deployments, including on-premises installations, are priced individually.

Arcade enters a market where Microsoft, Anthropic and OpenAI already offer ways to discover tools on demand. Its pitch combines that capability with a runtime that checks permissions, applies company policies and records agent activity across different models.

What Arcade already does, and who uses it

The first thing to understand is that Arcade is not a chatbot or a new large language model.

It supplies the behind-the-scenes software that allows AI agents powered by models such as Claude or ChatGPT to take actions in workplace systems.

For example, when an agent needs to update a Salesforce record or send an email through Gmail, Arcade can connect it to the application, check whether the employee and agent are authorized to take that action, handle the necessary credentials and record what happened.

The company's existing platform brings together three core capabilities:

• Security and permission checks for AI agents

• A catalog of ready-made software actions and tools for developers to build their own

• Central controls for administrators to manage and audit the actions agents take.

Its MCP Gateway can gather tools from Arcade and other providers behind a single connection. MCP, or Model Context Protocol, is an emerging common way of connecting AI assistants with outside applications. Arcade also launched a Private Registry on October 7, giving employees a place to see which tools their agents can use and request ones that are missing.

Arcade lists LangChain, Relevance AI, Snyk and Sybill among the organizations using its technology and publishes testimonials from representatives of those companies.

Its site also features Duracell and YUR, along with unnamed customers it describes as a Fortune 100 financial institution and a Nasdaq-100 consumer internet company. Those references do not demonstrate that any named customer has achieved the token savings reported for the new feature.

Co-founder and chief technology officer Sam Partee presented Arcade's security and oversight technology at the VentureBeat Transform 2026 Innovation Showcase in July.

At that event, he described how the platform lets agents act with the appropriate user permissions while keeping their activity visible to security teams.

The new Tool Recommendation system builds on that existing infrastructure: beyond determining whether an agent may act, Arcade now wants to make it easier and cheaper for the agent to identify what action to take.

Why an agent's tool list becomes a business expense

A tool is a callable function: search an inbox, retrieve a customer record, update a ticket or create a document. An agent needs descriptions of those functions and their required inputs before it can choose and use them.

According to Arcade's announcement, its catalog exceeds 8,000 tools. Describing the whole inventory would consume roughly 1.75 million tokens before a user submits a request. Even a much smaller catalog creates recurring overhead as an interaction continues.

Tool Recommendation replaces that inventory with a few entry points for discovering applications, requesting relevant functions and executing a selected action. The agent describes its immediate need; Arcade searches and ranks the catalog, returns a shortlist, and runs the chosen function.

Arcade uses smaller, specialized models to match and rank relevant tools instead of making a larger AI model repeatedly read the entire catalog. It does not replace the agent's underlying intelligence: The agent's own model still decides which suggested action to execute.

Partee told VentureBeat in an exclusive interview that Arcade builds on smaller models from the open-source ecosystem and data about actions taken by agents.

"We are able to take that data and train small language models to predict basically what you're going to want to do," he told VentureBeat in an interview.

Unlike a social media feed tailored to an individual, he said, many agent tasks fall into broader categories, such as sending an email, editing a draft or updating a Salesforce record.

Partee traced the work to an earlier site reliability engineering agent, when language models frequently struggled to select a tool and format its inputs.

"If you gave a model back then five tools, it wouldn't produce the right JSON," he recalled.

As models improved, the same recommendation approach could be applied more broadly to selecting actions and supplying the inputs needed to carry them out.

That could matter for work spanning several systems: A renewal process might require customer data, support history, financial information and a drafted response. Tool discovery allows different functions to enter the interaction when needed. Whether that improves a particular workflow depends on the quality of the shortlist and the cost of the extra searches.

Arcade says Tool Recommendation works through MCP and can be switched on at its gateway, the connection point between an agent and its available tools, without rebuilding existing integrations. The company says it can work across multiple models, agent-building frameworks and applications that support MCP. This matters because customers do not have to replace Claude, ChatGPT or their own agents to use the recommendation system.

The 37% claim and the measured 7.6% result describe different things

Arcade's methodology makes several distinctions that enterprise buyers should preserve.

The maximum 37% reduction comes from modeling input costs against an assistant supplied with 100 fixed tools. For a typical interaction containing three requests and approximately eight tool calls per request, Arcade estimates savings around 20%, depending on the model.

Partee offered a similar estimate in the interview: "You can just be your average user in a 10 turn session and be saving 18 to 20% of your tokens."

That figure describes an expected token reduction for a representative session, not independently verified savings in an enterprise customer's total bill.

The calculation excludes output tokens and assumes the client keeps large tool results outside the conversation. When those results remain in the interaction, the company estimates savings of only 5% to 7%. Short interactions produce the largest modeled percentage improvement.

Its projection of approximately $119,000 saved annually for 1,000 Claude Opus 5 users assumes 20 conversations per person per month. That is a scenario estimate, rather than reported savings from a named customer's deployment.

The 91% claim has a narrower denominator. It concerns input tokens associated with function descriptions and recommendation responses over a modeled 10-turn interaction, excluding the conversation itself and returned business data.

In that calculation, GPT-5.6's count falls from 256,650 to 22,164; Claude Opus 5's falls from 411,729 to 38,170.

"These are definition-token savings, not whole-session savings," Arcade writes in the methodology accompanying its announcement.

The more comprehensive test pairs workflows with and without discovery across 408 sessions covering 64 workflows, using Codex and Claude Code. Arcade reports a 7.6% reduction in total token charges, including input and output, with a 95% interval of 2.5% to 12.6% and unchanged task completion.

The reported reduction is 9.7% for Claude Code and 5.5% for Codex; the Codex interval includes zero. VentureBeat has not independently reproduced these tests.

Partee acknowledged that repeatedly calling the recommendation system could consume enough resources to erase the savings or even make an interaction more expensive. He maintained that this would be unusual but did not provide independently verifiable figures for how often it happens.

Its Microsoft comparison requires another qualification

Arcade also evaluates tool retrieval using ToolRet, a public benchmark. Recall@10 measures how many of the required tools appear among the first ten returned results. It does not measure completion of an entire business task.

Arcade supplied the following data to show how its Tool Recommendation system compares to an existing rival, Microsoft Foundry:

Sources: Arcade's announcement graphic and Microsoft's published evaluation. The runs use different settings.

Arcade's production setup includes an additional ranking stage and benchmark instructions. Microsoft's test omits the instruction string to avoid the expense and delay of generating it with another model call. Microsoft also describes a retrieval approach that avoids a resource-intensive ranking model during serving.

Arcade acknowledges that the comparison indicates a potential advantage rather than establishing a controlled head-to-head result. The scores warrant testing, but do not establish that Arcade delivers better task completion or lower total costs than Foundry.

Microsoft, Anthropic and OpenAI already address tool discovery

Microsoft's Foundry Toolboxes approach exposes search and execution entry points, then supplies the matching definitions. Microsoft describes it as independent of a particular model family and able to span different tool sources. Model flexibility therefore is not unique to Arcade.

Anthropic's Tool Search Tool similarly postpones loading definitions until Claude needs them. Its implementation supports built-in text-search methods as well as custom retrieval. Anthropic reports an 85% token reduction in its own example, but that uses a different baseline from Arcade's measurements.

OpenAI's Responses API supports deferred functions and MCP tools, with hosted or application-executed search. Its documentation says newly discovered definitions enter at the end of the context to preserve caching; support begins with GPT-5.4.

The distinction Arcade emphasizes is packaging discovery with execution controls in a layer customers can use across their chosen AI systems.

According to the company, that layer checks both a user's rights and the agent's permitted actions, keeps credentials outside the model's context, applies policy hooks before and after calls, and supplies audit records through OpenTelemetry. Those hooks can also support checks for sensitive data.

Partee said this approach also lets Arcade accommodate newer frontier models without rebuilding its recommendation system for each vendor.

After roughly three or four major model releases, he said, the technology's behavior had remained substantially consistent. "We should be able every time a new model is released to be able to guarantee something of the same level within a week," he said.

That is a forward-looking expectation, not an independently tested service guarantee; he also cautioned that the technology is changing quickly. He said the feature can work in Claude, ChatGPT and other agent environments.

Arcade's existing prices put the new savings claims in context

Arcade already charges for its underlying platform, with public prices based on usage rather than the number of employees. Its advertised options are:

Authorization events relate to connecting agents to user accounts; tool calls are the requests agents make to carry out specific actions. These published fees cover Arcade's platform, not the separate charges customers may incur for the large language models powering their agents.

Arcade says Tool Recommendation itself adds no separate feature charge for Arcade Cloud users. That should not be confused with unlimited free usage: Customers on the Team plan still face normal charges for authorization events and tool calls. Arcade also offers its underlying platform in customer-controlled clouds, on premises and in isolated environments. A separate enterprise rollout of the new recommendation feature is planned, but the company has not published a price or specific rollout schedule for those deployments.

Partee described larger organizations as the commercial focus, particularly those paying Arcade for security and management of agents across their workforces. For those customers, he said, "yes, I'm going to ask people to pay for that." His remarks concerned enterprise deployments, rather than the standard Arcade Cloud feature included at no extra charge.

That means a reduction in AI-model tokens does not automatically translate into the same percentage reduction in an enterprise's total AI spending. Buyers also have to account for Arcade's platform and per-action fees, any enterprise licensing costs and the cost of operating their own infrastructure. Arcade has not supplied a named customer's all-in cost comparison for the new feature.

The feature also is not automatically economical for every agent. Arcade's own modeling favors a fixed tool list below roughly 25 to 28 tools. For larger inventories, buyers need to measure search overhead, completion rates and the final bill on their own workloads.

For companies using several business applications and AI models, the practical proposition is broader than token savings: Arcade aims to provide one place to decide what an agent can do, connect it to the systems where work happens, record its actions and now suggest the relevant tool without cluttering the AI's conversation. The strongest evidence in its materials is the measured reduction in complete-workflow costs; the larger percentages explain where additional savings might arise under specific assumptions.