Give Your n8n AI Agent Tools (Calculator, HTTP, Workflows)

Your n8n AI Agent answers questions from training data alone, which means every answer could be months out of date. Attaching n8n AI agent tools fixes this by letting the agent call real functions (a live API, a calculation, a sub-workflow, your own document store) and act on current data instead of guessing.

That stale-data problem shows up as frustrated users, wrong answers, and an agent that can only chat. You end up patching prompts when the real fix is giving the agent the right functions to call.

What Does It Mean to Give an AI Agent Tools?#

A tool is a callable function, not a plugin or a prompt trick. The agent reads each tool's name and description, decides which to invoke, runs it, and feeds the result back into the model. This decide-call-read-repeat loop, documented in the n8n AI Agent reference, is what makes an agent different from a single LLM call.

Without tools, the model can only predict text. With tools, the same node can fetch a live stock price, submit a form, query your database, or fire off a Slack message, all in one agent run.

Adding tools changes what the agent can do and what it costs.

Which Tools Can You Attach in n8n?#

n8n's AI Agent node supports named tool sub-nodes you wire directly to its Tools connector. Each tool is a separate sub-node the agent can invoke during its loop, from a simple calculator to a whole sub-workflow. The n8n tools documentation lists the built-ins, and any HTTP or Code node can wrap a service that has no native tool.

Here is what each tool does in practice:

• HTTP Request Tool - calls any website or REST API. Use it to fetch live prices, hit your CRM, post to a webhook, or read a JSON feed.
• Custom Code Tool - runs a JavaScript or Python snippet the agent can call. Useful for data transformation or logic that does not fit a standard node.
• Call n8n Workflow Tool - wraps any existing n8n workflow as a callable function. Build a workflow once, reuse it as a tool across multiple agents. (Full reference for this sub-node.)
• Calculator - lets the agent do arithmetic reliably instead of asking the model to compute in its head (models are unreliable calculators).
• Wikipedia / SerpAPI - built-in tools for general knowledge lookups and web search.
• AI Agent Tool - lets one agent call another agent or sub-workflow, enabling multi-agent pipelines.
• Vector Store (retrieval mode) - set any Vector Store node to "Retrieve Documents (As Tool for AI Agent)" mode and it becomes a tool the agent can search on demand. That powers RAG-style customer support bots without hard-coding retrieval into every run. (See also: wiring a vector store for semantic search.)

The current default agent type in n8n is the Tools Agent, which is optimised for function calling. That is the one you want.

How Much Does an Agent with Tools Cost?#

More than you might expect. Each tool call is another full model round trip, and every round trip re-sends the system prompt, all tool definitions, and the conversation so far. A task that makes about 3 tool calls involves roughly 4 model round trips total.

Assuming approximately 6,500 input tokens and 600 output tokens across all round trips per task, here is what 1,000 agent runs costs in Claude tokens (as of June 2026, from the Anthropic pricing page):

The arithmetic: Haiku at $1.00 input and $5.00 output gives 6.5 x $1.00 + 0.6 x $5.00 = $9.50 per 1,000 tasks. That is several times the cost of a single classification or summarization call with the same model, purely because of the loop.

Running n8n self-hosted on a VPS (here is how to set that up for $0 in platform fees) keeps the platform cost at $0. n8n Cloud Starter is €20/month (n8n pricing page). The token cost is the variable you control most directly. Fewer tools, tighter descriptions, and lower max iterations all reduce it.

For a deeper look at controlling token spend on agentic workflows, see Claude API cost control for agent workflows.

Key figures for scoping a tool-using agent deployment.

How Does the Agent Decide Which Tool to Use?#

The model reads each tool's name and description on every round trip and picks whichever best matches what the task needs next. The description field is not a label, it is a prompt. "HTTP Request" tells the model almost nothing; "Fetch the EUR/USD rate from open.er-api.com, passing the base currency as a string" tells it exactly when and how.

When agents repeatedly skip the right tool, the fix is almost always rewriting that tool's description, not changing the model. The model is doing its job; it just lacks enough information to choose correctly.

Keep the tool list to 3 to 5 tools per agent. If you need more capability, build a second agent and connect the two with the AI Agent Tool sub-node.

How Do You Build It in n8n, Step by Step?#

Open your n8n canvas, add an AI Agent node, wire a Chat Model sub-node to its Chat Model connector, then add your chosen tool sub-nodes to its Tools connector. That three-part structure (agent + model + tools) is the complete wiring. The agent does not start until all three are connected and configured.

The minimum wiring to get a tool-using agent running.

Add the AI Agent node#

Click the canvas plus button, search AI Agent, and drag it onto the canvas. In the node settings, confirm the agent type is set to Tools Agent (the current default).

Attach a Claude Chat Model#

Click the Chat Model connector, add an Anthropic Chat Model sub-node, and enter your Anthropic API key. Select claude-haiku-4-5 for volume agents or claude-sonnet-4-6 when the reasoning needs to be more reliable. (For a tested comparison of Claude models in n8n, see Claude vs GPT vs Gemini in n8n agents.)

Wire in your tools#

Click the Tools connector. Add one sub-node per function:

1. HTTP Request Tool - set the method, URL, and any auth headers in the sub-node. The agent will call this when its description matches the task.
2. Calculator - no configuration needed. Add it and the agent can do arithmetic reliably.
3. Call n8n Workflow Tool - pick the target workflow from the dropdown. This turns any existing automation into a callable function.

Write the descriptions#

In each tool sub-node, fill the Description field with a plain-English sentence. Be specific about when to call it and what to pass. Vague descriptions are the single most common cause of agents picking the wrong tool.

Cap max iterations and test#

In the AI Agent node settings, set Max Iterations (5 to 10 is a safe starting range). Then send a real test query and open the execution log. Watch which tools the agent called and in what order. If it misses a tool, rewrite that tool's description.

Should You Add This as a Tool? (A Decision Guide)#

Before wiring a new capability as a tool, ask two questions. First: should the model choose when to invoke this, or should it always run? If it always runs, it belongs in a fixed workflow step, not a tool. Second: if the model calls this at the wrong time, what breaks?

Each new tool adds capability and cost. Weigh both before adding.

The decision tree above is the fastest gut-check. Anything involving sending messages, moving money, or deleting records belongs behind a human approval gate. For a full walkthrough of gating agent actions, see n8n AI human approval workflows.

Where Can This Go Wrong?#

Most agent failures are configuration failures, not model failures. The model is usually doing its job; the problem is how you set up the tools, the descriptions, and the loop. Four patterns cause the overwhelming majority of broken agents, and each has a fix you control. Here are the four that come up most often:

Too many tools. The model picks worse as the list grows. Each tool definition adds tokens to every round trip and introduces ambiguity. If your agent has more than 5 tools, split it into two agents.

Vague tool descriptions. A description like "calls the API" causes the agent to skip the tool or call it at the wrong time. Rewrite it with the exact trigger condition and the expected input format.

Runaway loops. A confused agent can loop many times, spending tokens on each pass. Always set a Max Iterations cap and track execution counts in your n8n logs.

Tools that fail silently. APIs time out; HTTP endpoints return 500s; workflows break. An agent that receives an error must handle it gracefully. Add error-handling logic to each tool sub-workflow and test what happens when the tool returns nothing useful.

High-stakes tools without gates. An agent that can send email or post to Slack autonomously will eventually do it at the wrong time. Keep those tools behind a human-in-the-loop approval. The n8n memory guide covers how to give the agent context across turns so it does not re-attempt a failed action on the next message.

What Should You Set Up This Weekend?#

Start with one tool, not five. Pick the capability your current agent most obviously lacks, probably live data via the HTTP Request Tool or reusable logic via the Call n8n Workflow Tool, and wire just that one. Write a specific description, cap iterations at 5, and run ten real test queries before adding anything else.

Once that single tool works reliably, add a second. After two tools are solid, you have a working pattern you can repeat. That is faster and cheaper than building a 10-tool agent that misfires unpredictably because several of its tool descriptions are ambiguous and you cannot tell which ones.

Check your execution logs after the first real-world day. If the agent is calling a tool zero times or every single time regardless of input, the description needs rewriting. Fix the description before assuming the model is wrong.