What is code mode for MCP?

Code mode lets an AI agent write code that calls its MCP tools as functions, instead of calling each tool itself. The name comes from Cloudflare's Code Mode. On camelRun every agent gets it by default: the model writes JavaScript in js_exec, where each MCP tool is tools. __ ().

Is code mode the same as programmatic tool calling?

They're the same idea: the model writes code that calls its tools, the calls run inside that code, and the model reads the code's result rather than every tool result. On camelRun that code is JavaScript, run in js_exec.

Why let the model write code to call tools?

Code can loop, filter and call several tools in one step. The model writes the job once, the calls run together, and it reads back the result instead of every row along the way. A large tool catalog also stays out of the prompt until the code asks for a tool.

How does MCP tool search work?

The agent's code calls tools.search with a few words and gets the best matches first, each with its name, description and input. tools.namespaces lists the servers with their tool counts. On hosted camelRun the ranking uses meaning as well as keywords.

Does code mode work with my MCP servers?

Yes. Connect your MCP servers in the agent's definition or in its own server list, and their tools can be called from code. Servers with up to 10 tools are offered both ways by default and bigger ones from code only. Set exposure on the server entry, or on a single tool in its tools/list, to change that.

Which tools does the model call directly?

Every tool you mark direct, and every tool that needs a person's approval. An agent made with codeMode: false calls all of its tools directly, which suits an agent that answers in a line, such as a classifier.

What language does the model write in code mode?

JavaScript. Your tools stay in your own app, in TypeScript or Python with camelRun's SDKs, and the model's code calls them by name: one at a time, in a loop, or together with Promise.all.

Is it safe to let the model run code?

Yes. The model's code runs in a sandbox that can reach nothing else: it can call the agent's tools and work with its files, and your tools run in your own app.

Code mode · MCP · Tool search

Code mode for AI agents

camelRun by camelAI hosts AI agents that call tools from code: one step, many calls, one result.camelRun by camelAI hosts AI agents that call their tools from code. The model writes a few lines of JavaScript that filter and call many tools in one step, your MCP servers included, and reads back the script's result.

Get an API key Quickstart

camelRun runs: model turns and js_execYour tools

Agentinventory

completed

You

Plan restocks for everything below target.

js_execthe model's code

                    const rows = await tools.read_inventory({});
                    const low = rows.filter(row => row.stock < row.target);
                    return await Promise.all(low.map(row => tools.plan_restock({sku: row.sku, quantity: row.target - row.stock})));

Code mode

model

js_exec

read_inventory

plan_restock

BEAN-01→ 32

COCO-08→ 15

CROI-11→ 20

DECAF-12→ 9

FILT-13→ 80

MILK-05→ 27

OAT-02→ 24

SUG-09→ 40

SYRP-06→ 10

TEA-07→ 15

model

model reads

Direct tool calls

model

your tools

model reads

camelRun runs: model turns and js_execYour tools

How it works

The model writes JavaScript that calls your tools

Your tools stay in your app. Here they're two Python functions over a cafe's stock list in SQLite. In js_exec, each tool is a function the model can call in a loop, after a filter, or all at once.

01
Loops and filters run in code
The script reads all fourteen rows and keeps the ten below target. The model gets the script's result, not every row.
02
The calls go out together
Promise.all sends all ten restocks at once, inside one step of the agent's turn.
03
Calls from code are on the record
The run lists the calls the script made, each with js_exec's toolCallId and an innerCallId of its own.

run · toolCalls

11 calls

js_exectoolCallIdtoolu_01M4kQ8vZ2cWn7RtYb3xJp

tool	innerCallId	ok
read_inventory	toolu_01M4kQ8vZ2cWn7RtYb3xJp…:1	true
plan_restock	toolu_01M4kQ8vZ2cWn7RtYb3xJp…:2	true
plan_restock	toolu_01M4kQ8vZ2cWn7RtYb3xJp…:3	true
… :4 to :10
plan_restock	toolu_01M4kQ8vZ2cWn7RtYb3xJp…:11	true

inventory.py

GitHub

14from camelai_run import Agents, ToolContext, tool

28    @tool
29    async def read_inventory():
30        """Read inventory and target stock from the application's SQLite database."""
31        return [dict(row) for row in db.execute("SELECT * FROM inventory ORDER BY sku")]
32
33    @tool
34    def plan_restock(sku: str, quantity: int, context: ToolContext):
35        """Set one SKU's local restock quantity. Does not place orders."""
36        if type(quantity) is not int or not 1 <= quantity <= 100:
37            raise ValueError("Invalid restock quantity")
38        if not db.execute("SELECT 1 FROM inventory WHERE sku = ?", (sku,)).fetchone():
39            raise ValueError("Unknown SKU")
40        # The call's stable key: a retried call rewrites the same row.
41        db.execute("INSERT OR REPLACE INTO restock VALUES (?, ?, ?)", (sku, quantity, context.idempotency_key))
42        db.commit()
43        return {"planned": True, "sku": sku, "quantity": quantity}
44
45    async with Agents() as agents:
46        agent = await agents.upsert("inventory", instructions="You manage cafe stock. Plan restocks without purchasing anything.",
47                                    tools=[read_inventory, plan_restock], **({"model": os.environ["AGENT_MODEL"]} if os.environ.get("AGENT_MODEL") else {}))

Code mode MCP

Code mode for MCP servers

Each tool on an MCP server you add is a function the agent's code can call, named tools.<server>__<tool>, and a big catalog leaves the model's context free for the job.

You decide, per server or per tool, and the console's agent page shows which way each tool is offered.

direct: the model calls it as a tool of its own; never from js_exec
codemode: only from code the model runs in js_exec
both: as a tool of its own, or from code in js_exec

Console/Agents/inventory

Tools the model gets (117)

Tools marked codemode or both can be called from code the model runs in js_exec. To keep one out of js_exec, set its exposure to "direct": on the tool in your code, or on the server in its definition. How

Your applicationconnected

read_inventoryboth

Read inventory and target stock from the application's SQLite database.

plan_restockboth

Set one SKU's local restock quantity. Does not place orders.

MCP serversupplierhttps://mcp.example.com/supplierlisted 09:41

supplier__search_productscodemode

Find products in the supplier's catalog by name or SKU.

supplier__get_pricecodemode

Price per case for one product, with volume breaks.

supplier__get_lead_timecodemode

Days from order to delivery for one product.

supplier__list_delivery_slotscodemode

Open delivery slots for the next two weeks.

MCP serverposhttps://mcp.example.com/poslisted 09:41

pos__list_salescodemode

Sales for one day, item by item.

pos__top_itemscodemode

Best sellers between two dates.

Tool search

MCP tool search for agents with many tools

Give one agent the servers for its whole job. Its code finds the right tool with tools.search, reads the schema with tools.describe, and calls it in the same script.

Your app 2

pos 38

rota 24

supplier 46

js_exec

best matches first

await tools.search("price per case for oat milk")

1
name: "supplier__get_price"
description: "Price per case for one product, with volume breaks."
input: "{ sku: string }"
2
name: "supplier__list_price_breaks"
description: "Case prices by quantity for one product."
input: "{ sku: string }"
3
name: "supplier__get_case_size"
description: "Units in one case of a product."
input: "{ sku: string }"

Thousands of tools, nothing up front: Only tool names enter the code's environment, so a catalog of thousands stays out of the script's way until it asks.
Best matches first: Each result comes back with its name, description and input, the most relevant at the top.
Ranked by meaning: On hosted camelRun, search ranks by what a tool does as well as by its keywords.

Test it yourself

Run the model's code yourself to test your tools

agent.client.execute runs a script with the agent's tools, outside the model's history. Write the code the model would write, run it from your app, and see what your tools did.

inventory.py

56                # Code the model could have written, run in the agent's sandbox with its tools (no model call).
57                result = await agent.client.execute("""
58                    const rows = await tools.read_inventory({});
59                    const low = rows.filter(row => row.stock < row.target);
60                    return await Promise.all(low.map(row => tools.plan_restock({sku: row.sku, quantity: row.target - row.stock})));
61                """)

Your databaseSELECT sku, quantity FROM restock ORDER BY sku

The restock table after the code ran
sku	quantity	sku	quantity
BEAN-01	32	MILK-05	27
COCO-08	15	OAT-02	24
CROI-11	20	SUG-09	40
DECAF-12	9	SYRP-06	10
FILT-13	80	TEA-07	15

FAQ

Code mode and tool search questions

q1What is code mode for MCP?: Code mode lets an AI agent write code that calls its MCP tools as functions, instead of calling each tool itself. The name comes from Cloudflare's Code Mode. On camelRun every agent gets it by default: the model writes JavaScript in js_exec, where each MCP tool is tools.<server>__<tool>().
q2Is code mode the same as programmatic tool calling?: They're the same idea: the model writes code that calls its tools, the calls run inside that code, and the model reads the code's result rather than every tool result. On camelRun that code is JavaScript, run in js_exec.
q3Why let the model write code to call tools?: Code can loop, filter and call several tools in one step. The model writes the job once, the calls run together, and it reads back the result instead of every row along the way. A large tool catalog also stays out of the prompt until the code asks for a tool.
q4How does MCP tool search work?: The agent's code calls tools.search with a few words and gets the best matches first, each with its name, description and input. tools.namespaces lists the servers with their tool counts. On hosted camelRun the ranking uses meaning as well as keywords.
q5Does code mode work with my MCP servers?: Yes. Connect your MCP servers in the agent's definition or in its own server list, and their tools can be called from code. Servers with up to 10 tools are offered both ways by default and bigger ones from code only. Set exposure on the server entry, or on a single tool in its tools/list, to change that.
q6Which tools does the model call directly?: Every tool you mark direct, and every tool that needs a person's approval. An agent made with codeMode: false calls all of its tools directly, which suits an agent that answers in a line, such as a classifier.
q7What language does the model write in code mode?: JavaScript. Your tools stay in your own app, in TypeScript or Python with camelRun's SDKs, and the model's code calls them by name: one at a time, in a loop, or together with Promise.all.
q8Is it safe to let the model run code?: Yes. The model's code runs in a sandbox that can reach nothing else: it can call the agent's tools and work with its files, and your tools run in your own app.

Build an agent with code mode

Every agent on camelRun can write code from its first turn. Get an API key, hand it your tools, and give it a job.

$0.01 per agent-hour while an agent is in a run. Tokens on your own provider keys cost $0, and on ours the provider's price, no markup on tokens.

Get an API key Quickstart Tools docs

Set up with your coding agent

Read https://run.camelai.com/SKILL.md and set up camelRun in this project. Ask me what the agent should do.

Paste into Claude Code, Codex, Cursor or any coding agent. It asks what your agent should do, then builds it.

Code mode for AI agents

The model writes JavaScript that calls your tools

Loops and filters run in code

The calls go out together

Calls from code are on the record

Code mode for MCP servers

MCP tool search for agents with many tools

Run the model's code yourself to test your tools

Code mode and tool search questions

More on tools for AI agents

Build an agent with code mode