SSnowflake CoCo (Cortex Code)

Snowflake CoCo Writes Data Code With the Account in View

TS Imagine’s data engineers submitted five times more pull requests after adopting Cortex Code, Snowflake’s coding agent, according to Snowflake’s March 26, 2026 announcement that the agent was generally available in Snowsight. Snowflake first launched the agent on February 3, 2026 as a CLI, and its documentation now calls the product Snowflake CoCo.

CoCo suits data teams on Snowflake who want an agent that knows their tables, roles and compute while it writes SQL, Python, dbt and Streamlit code. Snowflake accounts pay for token use as they go, and individuals get a 30-day trial of the CLI and Desktop app. CoCo scores 8.9/10 on our scale.

CoCo Launch Dates, Customers and Usage Results

The launch release, the general availability blog, Snowflake’s docs and G2 supply these figures.

The TS Imagine figure is the only hard usage number Snowflake has published for CoCo so far. Five times more pull requests says the agent sped up output. It does not say whether review time or defects changed, which is what a data lead will want to know next.

The launch customer list spans marketing tech, advertising, lending, insurance and equipment rental. WHOOP used the agent to tune its own Cortex Agents and score them against evaluation sets, an early sign that teams use CoCo to build other agents as well as pipelines.

The gap of about seven weeks between the CLI launch and Snowsight availability also tells buyers something. Snowflake shipped the developer surface first, then the browser version every user can reach, then a desktop app. Each step widened the audience, from engineers in a terminal to analysts and admins in the browser.

How CoCo Works Across Snowsight, Desktop and the CLI

CoCo has one model of the account and three places to use it. Each suits a different kind of user.

CoCo Inside Snowsight

In Snowsight, CoCo sits inside Workspaces and admin pages. It writes SQL and Python notebooks, answers admin questions about credits, query performance, governance and permissions, and shows a diff before applying a change. Admins turn it on from the CoCo Admin page. Snowflake says every Snowflake user has access in Snowsight with no extra setup.

CoCo Desktop

CoCo Desktop is a standalone IDE for macOS and Windows. It has an agent mode and an editor mode, reads local files and supports plugins, skills, hooks and MCP. It also includes an agentic browser, so the agent can look things up on the web while it works.

CoCo CLI

The CLI connects a developer’s terminal to Snowflake. It can run SQL, read and write local files and use git. It follows AGENTS.md files, supports Agent Skills and MCP, works with editors through the Agent Client Protocol and respects Snowflake role-based access control. A three-tier approval system and sandboxing control what it may do without asking. The CLI runs natively on Windows through PowerShell as well as on macOS and Linux.

Agent Teams and Built-In Skills

Agent Teams let a lead agent split a job across parallel subagents that share a task list. Skills are packaged instructions that load when needed. Snowflake ships skills for Streamlit apps, Openflow, cost intelligence, machine learning and Cortex AI functions. Teams can stand up dbt pipelines in Snowsight, and the CLI supports Apache Airflow.

A dbt Model Change, Step by Step

An analytics engineer asks CoCo in the CLI to add a revenue-by-region model to the team’s dbt project. CoCo reads AGENTS.md for the team’s conventions, checks which tables the engineer’s role can read and drafts the model. It asks approval before running dbt against the development warehouse, reads the test results and fixes a failing join. The engineer reviews the diff, commits it with git and opens a pull request.

Turning On CoCo for an Account

An account admin enables CoCo from the CoCo Admin page in Snowsight. Each surface lists the account types it supports. All three surfaces need cross-region inference turned on, which lets Snowflake process model requests outside the account’s home region. Once enabled, users reach CoCo inside Workspaces, and developers install the CLI or Desktop app and sign in to the account.

Governance Questions to Settle Before Rollout

Snowflake’s docs classify what users type into CoCo as usage data and what CoCo produces as customer data. Model use falls under Snowflake’s model and service pass-through terms, which legal teams should read before approval.

On the control side, the CLI respects Snowflake role-based access control, runs in a sandbox and uses a three-tier approval system for actions. Snowsight shows a diff before any change applies. The cross-region inference requirement is the point most likely to need sign-off, since it means prompts may be processed in another region. A short written policy on which roles may use CoCo, and on which warehouses, will settle most reviews.

What CoCo Costs on a Snowflake Bill

Snowflake bills CoCo by tokens. The amount a team pays depends on the surface and the type of account.

Planning a CoCo Budget for a Data Team

A team pays twice for agent work: tokens for the model and credits for the warehouse queries the agent runs. A query the agent runs costs the same compute as one a person runs, so the second line grows with how much the agent executes.

Snowflake offers three controls. Admins can set daily credit usage limits for CoCo, cost controls cap spending, and the CORTEX_CODE_CLI_USAGE_HISTORY view in Account Usage records CLI usage for reporting. The built-in cost-intelligence skill can also answer questions about credit use. We suggest setting a daily limit before a team-wide rollout, then reviewing the usage view weekly for the first month.

Three habits drive most of the bill. Long agent sessions that keep large tables in context use more tokens. Agent Teams multiply token use, since each subagent works in parallel. Agents that rerun failing queries add warehouse credits each time. Keeping requests narrow and pointing CoCo at a small development warehouse keeps both lines down.

Snowflake could make planning clearer for buyers by publishing the individual subscription price and per-token rates on the main CoCo page.

CoCo Compared With Databricks Genie Code, Hex and Julius AI

Two of these agents belong to data platforms, and two are separate analysis tools.

CoCo and Genie Code mirror each other. Each knows its own platform’s permissions and compute, and each works best when the data already lives there. CoCo’s extra reach is the desktop IDE and a CLI that works inside VS Code or Cursor. Hex and Julius AI cost a flat monthly fee and connect to several sources. More tools sit in our data analysis agents list.

Picking a CoCo Surface for Each Role

Analysts and admins who already live in the browser should start in Snowsight. It needs no install, shows diffs before changes and covers notebooks, credits, query performance and permissions.

Analytics engineers with dbt projects in git will get more from the CLI. It reads AGENTS.md, works in VS Code or Cursor, commits with git and supports Airflow. Its three-tier approvals let a team decide which actions run unattended.

Developers building Streamlit apps or agents may prefer CoCo Desktop, which adds an editor mode, hooks, plugins and an agentic browser for looking up documentation. Teams that want to embed CoCo in their own tools can watch the Agent SDK, still in preview for Python and TypeScript.

Snowflake Teams That Gain Most From CoCo

Analytics engineers who maintain dbt projects gain from an agent that reads their conventions, checks permissions and runs tests. Data engineers who build Openflow or Airflow pipelines can hand off first drafts and debugging.

Teams building their own agents on Snowflake also fit. WHOOP used CoCo to tune Cortex Agents against evaluation sets. Small data teams with one or two engineers may get the most relative lift, since an agent that drafts models and debugs pipelines adds capacity they cannot hire quickly. Snowflake admins can ask CoCo about credit use, slow queries and permissions in Snowsight without writing the account-usage SQL by hand.

Who May Hold Off on CoCo for Now

Companies whose policies forbid cross-region inference cannot turn CoCo on, since every surface requires it. They should watch for changes to that requirement.

Teams that need published per-token prices before approving a tool may also wait, or ask Snowflake sales for a written rate card first.

Independent developers on the individual plan should know that the tool stops working once the monthly usage amount is spent, until the next billing period. Teams that want to build on the Agent SDK may wait until it leaves preview.

Cost and Scale Feedback in Snowflake’s G2 Reviews

Snowflake holds 4.6 stars from 762 G2 reviews of the platform. The top praise themes are ease of use, with 183 mentions, a reliable feature set, data integration, scaling and fast processing.

Cost dominates the complaints. High costs draw 91 mentions, cost control issues 44 and a hard-to-optimize cost structure 44. Since agent-run queries add compute, CoCo users will want daily limits set early. These reviews cover Snowflake broadly, and CoCo-specific reviews are not yet broken out.

CoCo From CLI Launch to Desktop App

• February 3, 2026: Snowflake launches Cortex Code with a generally available CLI.
• March 26, 2026: Cortex Code in Snowsight becomes generally available, with Agent Teams and Windows CLI support.
• 2026: Snowflake’s docs rename the product CoCo and add CoCo Desktop and a preview Agent SDK.

How Snowflake CoCo Scores for Snowflake Data Teams

CoCo scores 8.9/10. It scores well on autonomy, integrations and trust: it runs code across three surfaces, coordinates subagents and works under Snowflake role-based access with approvals and sandboxing. Pricing value scores lower since token rates and the individual subscription price are not on the main docs page, and compute adds a second charge.

For a team already on Snowflake, CoCo is the most direct way to add a coding agent that knows the account. Set daily credit limits, start with dbt or notebook work and review the usage history before opening it to every user.

What Snowflake Users Ask About CoCo

Is Cortex Code the Same Product as Snowflake CoCo?

Yes. Snowflake launched it as Cortex Code in February 2026, and its documentation now calls it Snowflake CoCo.

Can I Use CoCo Without a Snowflake Account?

Individuals can try the CLI and Desktop free for 30 days, then pay a subscription with a fixed monthly usage amount. The agent connects to a Snowflake account to work with data.

How Is CoCo Billed?

By token consumption. Snowflake accounts pay as they go, and compute and storage are billed separately at on-demand rates. Admins can set daily credit usage limits.

Does CoCo Respect Snowflake Roles?

Yes. The CLI supports Snowflake role-based access control, and the agent works within the permissions of the user who runs it.

Can CoCo Run Several Tasks at Once?

Yes. Agent Teams let a lead agent coordinate parallel subagents through a shared task list, so one request can be split into several pieces of work.

Does CoCo Need Cross-Region Inference?

Yes. Snowsight, Desktop and CLI all require cross-region inference to be enabled on the account.

What Is CoCo Desktop?

A standalone IDE for macOS and Windows with agent and editor modes, local file access, plugins, skills, hooks, MCP support and an agentic browser.

Which Editors Work With the CoCo CLI?

Snowflake names VS Code and Cursor, and the CLI supports the Agent Client Protocol for other editors.