TL;DR
Snowflake CoCo is Snowflake’s built-in AI coding agent, renamed from Cortex Code at Summit 2026. It reads a team’s actual schemas, RBAC permissions, and lineage before generating SQL, dbt models, Airflow DAGs, or ML pipelines, then executes the work inside Snowsight, a desktop IDE, or a CLI. On Snowflake’s own ADE-Bench testing, CoCo scored 72.1 percent accuracy against 65.1 percent for Claude Code and OpenAI Codex , though that figure has not been independently replicated. Both Snowsight and CLI usage now consume token-based AI Credits after an earlier free preview ended. This guide covers how CoCo works, what it costs, how it compares to Claude Code and GitHub Copilot, and where it fits an enterprise data strategy.
Data teams have spent the past two years watching general-purpose coding agents write plausible SQL against warehouses they know nothing about. The queries compile, but the joins are wrong, the RBAC gets ignored, and someone downstream inherits the cleanup.
Snowflake built CoCo to close that gap by reading a team’s actual schemas, roles, and lineage before it writes a line of code. The tool spent its first few months under the name Cortex Code before Snowflake renamed it at Summit 2026.
This guide walks through how CoCo works, what it costs, how it compares to Claude Code and GitHub Copilot, and where it fits inside a broader data strategy.
Key Takeaways Snowflake CoCo is the Summit 2026 rebrand of Cortex Code, with the same underlying product and architecture. It generates code grounded in real schemas, RBAC roles, and lineage rather than generic pattern matching. CoCo runs in Snowsight, a native desktop app, a CLI, and inside VS Code or Claude Code through official plugins. Both Snowsight and CLI sessions now bill against token-based AI Credits, ending the earlier free preview period. Security teams should treat CoCo’s shell and file access the same way they treat any agent with local credentials.
What Is Snowflake CoCo Snowflake CoCo is Snowflake’s native AI coding agent for data and AI teams. It translates natural language requests into production-ready SQL, dbt models, Airflow DAGs, and machine learning pipelines, then plans, reviews, and executes the work instead of only suggesting snippets. The agent reads an account’s real catalog , lineage, and role-based access controls before it writes anything, which Snowflake positions as the core difference versus generic coding assistants pointed at a database from the outside.
1. From Cortex Code to CoCo The product launched as Cortex Code in February 2026 and operated under that name for several months before Snowflake renamed it CoCo at Snowflake Summit 2026 on June 2. The rename did not touch the underlying architecture or capability set. Anyone who built workflows around Cortex Code keeps the same behavior under the new name.
2. How CoCo Understands Your Snowflake Environment CoCo checks which tables actually contain the data a request needs, verifies read access under the caller’s role, and confirms upstream freshness before drafting a plan. A generic coding agent pointed at a dbt repository has no visibility into any of that context, so a developer has to supply it by hand. That grounding step is what separates a data-native agent from a chat window bolted onto a code editor.
Kanerika’s Snowflake Consulting and Migration Services Kanerika is a Snowflake Select Tier Partner that builds the governed, well-modeled Snowflake environments that make tools like CoCo generate accurate code in the first place.
Explore Snowflake Services →
How Snowflake CoCo Works Across Its Tools and Integrations CoCo ships in three primary surfaces, and all three read from the same underlying context layer. Which one a team reaches for depends on whether the work is exploratory analysis, day-to-day pipeline development, or automation that needs to run unattended.
1. Snowsight and Workspaces Inside Snowsight , CoCo sits alongside SQL editors, notebooks, and the admin console. It answers questions about credit consumption, query performance, and governance settings directly, and it can generate or explain SQL and Python notebook code in place. This is the entry point most analysts and less technical users reach first, since it needs no local installation.
2. CoCo Desktop CoCo Desktop is a native macOS and Windows application Snowflake introduced at Summit 2026, combining a file browser, integrated terminal, and full editor with the same agentic loop available elsewhere. It keeps project context across sessions, so a developer does not have to re-explain schema decisions or prior work each time they reopen the app. Cloud Agents, a related Summit 2026 feature, let long-running jobs such as a multi-hour dbt build finish in a managed container even after a laptop goes to sleep.
3. CoCo CLI and IDE Integrations The CLI runs in a local terminal and bridges a developer’s existing tools, VS Code , Cursor , or a plain shell, with a live Snowflake connection. Snowflake also ships a dedicated CoCo extension for VS Code and a CoCo plugin for Claude Code , so teams already standardized on another agent can add Snowflake awareness without switching editors. The CLI also supports the Model Context Protocol and, in preview, the Agent Client Protocol, letting editors such as Zed or JetBrains drive a CoCo session directly.
Surface Best For Local File Access Runs Unattended Snowsight SQL, notebooks, admin questions No No CoCo Desktop Full agentic development outside the browser Yes Cloud Agents only CoCo CLI (plus the VS Code extension and Claude Code plugin) Terminal-based dbt and pipeline work, CI, or working from inside an existing editor Yes Yes, via Automations
What Snowflake CoCo Can Do Inside a Data Stack CoCo organizes its work around Skills, prebuilt and specialized routines for common data engineering tasks. A team does not need to invoke a Skill by name. Describing the outcome in plain language is enough for CoCo to pick the right combination of tools automatically.
1. Data Engineering and dbt Workflows CoCo drafts dbt models, Airflow DAGs, and Streamlit apps from a plain-language description of the desired outcome, then checks the draft against real table structures before handing it back for review. A detailed CoCo capabilities breakdown describes one team building a custom Skill for dbt performance tuning that cut a model’s runtime from roughly 10 hours to under 2, then shared it across the organization as a reusable workflow.
2. Cloud Agents for Long-Running Jobs Cloud Agents run inside isolated managed containers within Snowsight, supporting shell commands, Python scripts, package installation, and dbt builds and tests. A multi-hour training run or a large dbt build finishes in the cloud and reports back, so a developer’s machine does not need to stay open the entire time.
3. Automations and Scheduled Workflows Automations are recurring, event-driven workflows that fire on a schedule or in response to a trigger such as new data arriving, a threshold breach, or a pipeline failure. Snowflake documents examples including daily pipeline health checks and weekly model drift detection. Automations run under the identity of the user who created them, not a shared service account, and inherit that user’s role-based access controls automatically.
Snowflake CoCo Pricing and the AI Credit Model Cost is the question every finance and platform team asks once a pilot goes well. The answer changed materially in early 2026, and teams still pricing from the original announcement are working from stale numbers.
1. How Snowsight and CLI Billing Works Snowsight access was free during CoCo’s initial preview window, but that period ended in early 2026 . Both Snowsight and CLI usage now consume token-based AI Credits at the rates published in Snowflake’s consumption table. CoCo also triggers standard Snowflake Cloud Services compute whenever it queries metadata, so token charges are not the only cost line to watch.
2. Why Session Length Affects Cost Cortex Code charges per conversation turn, and starting with the second turn, repeated context such as conversation history and file contents gets served from cache at roughly a tenth of the normal input price. That pricing structure means a handful of long, focused sessions typically costs less than the same work spread across many short, disconnected ones.
3. Tracking and Attributing CoCo Spend CoCo costs roll up into the broader AI Services bucket inside Snowsight Cost Management and the account’s METERING_DAILY_HISTORY view, which gives a top-line number but not enough detail for chargeback. Per-user, per-session detail lives in two dedicated views, CORTEX_CODE_SNOWSIGHT_USAGE_HISTORY and CORTEX_CODE_CLI_USAGE_HISTORY, both exposing token and credit counts at the individual turn level. Joining either view against SNOWFLAKE.ACCOUNT_USAGE.USERS turns raw user IDs into names a finance team can actually act on.
4. Setting Per-User Credit Limits Account admins can cap daily spend independently for each surface using CORTEX_CODE_CLI_DAILY_EST_CREDIT_LIMIT_PER_USER and CORTEX_CODE_SNOWSIGHT_DAILY_EST_CREDIT_LIMIT_PER_USER, set at the account level or overridden for individual users. The two limits track separately, so a user who hits the CLI ceiling keeps full Snowsight access and vice versa. Treat these as a backstop against runaway usage, not a substitute for actually reviewing which sessions are burning the most credit.
Snowflake CoCo vs Claude Code, GitHub Copilot, and Cursor CoCo is not trying to replace a developer’s primary editor. Snowflake’s own documentation recommends pairing CoCo with an existing tool for general editing, then using CoCo specifically for the Snowflake-aware parts of the job.
1. Performance on ADE-Bench Snowflake reports that CoCo scored 72.1 percent on ADE-Bench , a benchmark dbt Labs built specifically for analytics engineering tasks, against 65.1 percent for both Anthropic’s Claude Code and OpenAI’s Codex. The benchmark itself is maintained independently by dbt Labs, but the CoCo score is Snowflake’s own measurement and has not yet been replicated by an outside party, which matters more than the seven-point gap when deciding how much weight to put on the result. The comparison also measures a narrow task type, analytics engineering work with known schemas and governance context already in place, where a data-native agent has a structural advantage by design and says less about how CoCo performs on general software engineering tasks outside a Snowflake account.
2. Where Each Tool’s Strength Actually Sits The table below breaks down what each tool is built for rather than which one scores higher on a single benchmark.
Tool Primary Strength Snowflake Context Awareness Best Use Case Snowflake CoCo Native schema, RBAC, and lineage grounding Deep, account-specific by default Analytics engineering, dbt, and pipeline automation inside Snowflake Claude Code General-purpose software engineering across any codebase None by default, added via the CoCo plugin Full-repository development, refactors, cross-stack projects GitHub Copilot IDE-integrated autocomplete and chat, wide language coverage None (Fabric-specific awareness is a separate product) Everyday coding inside an existing IDE workflow Cursor AI-native code editor with repo-wide context None Teams that want an AI-first editor rather than an add-on agent
Kanerika’s side-by-side comparison of GitHub Copilot, Claude Code, Cursor, and Windsurf breaks down pricing and benchmarks across those four tools in more depth for teams evaluating options outside Snowflake specifically. Most enterprise teams end up running two tools rather than one, an AI coding agent built for general engineering plus CoCo for the Snowflake-specific work neither Claude Code nor Copilot can see into on its own.
3. How Pricing Models Differ Across the Four Tools CoCo is the outlier on pricing structure. It bills purely on token-based AI Credits with no flat subscription tier, so cost tracks usage directly instead of a fixed monthly fee, while Claude Code and GitHub Copilot both ship inside flat-rate subscription plans, Anthropic’s Pro and Max tiers for the former and a monthly or annual seat license for the latter, with usage limits rather than per-token billing. Cursor mixes the two models, a subscription tier plus metered overage once included usage runs out, which puts a team running CoCo alongside any of the other three in the position of paying a flat editor fee plus variable Snowflake consumption on top.
4. When Running Two Agents Makes Sense Pairing CoCo with a general-purpose agent is not redundancy, it covers two different blind spots. CoCo sees everything inside a Snowflake account and nothing outside it, while Claude Code, Copilot, and Cursor see an entire repository but nothing about RBAC, lineage, or governed schemas unless a plugin bridges the gap. Teams that skip CoCo end up asking a general agent to guess at Snowflake context it cannot verify, and teams that skip a general agent lose coverage for everything CoCo cannot touch outside the warehouse.
Security and Governance Considerations for Snowflake CoCo CoCo’s governance story is stronger than a typical AI coding tool because it operates inside Snowflake’s existing permission model rather than a separate one. That does not make it risk-free, and security teams evaluating it should look past the marketing to the actual attack surface.
1. RBAC and Row-Level Security CoCo respects row-level security, masking policies, and every governance control already configured in an account, which Snowflake positions as suitable for regulated industries including financial services , healthcare, and insurance. Two database roles gate access. SNOWFLAKE.COPILOT_USER is required for every user, and either SNOWFLAKE.CORTEX_USER or SNOWFLAKE.CORTEX_AGENT_USER is required depending on how advanced the intended workflows are.
2. Credential and Shell Access Risks The CLI runs with whatever permissions launched it, so a shell environment holding AWS credentials, SSH keys, or a Snowflake password in plain text is readable by the agent too. A security review from 7Rivers found this exposure real enough to warrant containerizing the CLI with Docker rather than relying on the default sandbox alone. Every tool call also runs under a tiered risk model that decides whether the CLI prompts for confirmation first, though that classification is only as trustworthy as the environment it runs in.
3. Network Egress and Containerization Unrestricted network access through the CLI’s web tool and shell commands is the second major exposure alongside credential handling. Without egress controls in place, a compromised session, or a prompt injection hidden inside a file the agent reads, has no real barrier stopping it from reaching an external endpoint. Docker containerization with explicit network policies closes that gap more reliably than the CLI’s default sandbox, which the 7Rivers review found insufficient for regulated environments on its own.
4. Deployment and Account Restrictions CoCo is unavailable on Government, VPS, and Sovereign Snowflake deployments regardless of governance configuration, and the account needs cross-region inference enabled before anyone can turn it on. Accounts that previously opted out of the legacy Snowflake Copilot have CoCo disabled by default too, so re-enabling it means a conversation with the Snowflake account team rather than a self-service toggle. Security and platform teams evaluating CoCo for a regulated deployment should confirm account eligibility before promising it to end users.
Snowflake CoCo: How Kanerika Helps Data Teams Get More From It Kanerika is a Microsoft Solutions Partner for Data and AI with Analytics Specialization and Microsoft Fabric Featured Partner, and holds Snowflake Select Tier Partner status alongside its Databricks Consulting Partner credential. The firm has spent over 10 years building data platforms for more than 100 enterprise clients, with a Snowflake practice that sits inside a broader multi-platform footprint rather than a single-vendor playbook. That comparative delivery experience across Snowflake , Databricks , and Fabric shapes how Kanerika advises clients evaluating a native tool like CoCo against a wider modernization plan .
Snowflake CoCo speeds up work inside an account that is already well-modeled, but it cannot fix a Snowflake environment built on undocumented schemas, orphaned tables, or reconciliation logic that still runs by hand. Kanerika’s Snowflake migration and modernization work focuses on that foundation first, using the FLIP accelerator to automate pipeline conversion and validation, which cuts migration effort by 50 to 60 percent and typically completes in 2 to 8 weeks depending on pipeline count. A cleaner, governed data estate is what lets a tool like CoCo generate accurate code in the first place.
Kanerika also builds and deploys named AI agents on top of platforms like Snowflake, including Karl for natural-language analytics, both of which inherit a client’s existing governance and access controls rather than operating outside them. Teams weighing CoCo against a broader AI agent strategy get a second opinion grounded in production deployments, not vendor materials. Kanerika’s Snowflake partnership page documents the credential and service scope in full.
Transform Your Business with AI-Powered Solutions! Partner with Kanerika for Expert AI implementation Services
Book a Meeting
Case Study: How Kanerika Cut Manual Reconciliation With a Snowflake Migration Challenges Regional systems operated independently, so reconciling data across them required manual, spreadsheet-driven work every reporting cycle. Distributed teams only saw operational visibility in month-end reports, well after decisions needed to be made. The reconciliation logic itself was undocumented and depended on institutional knowledge held by a small group of analysts.
Solutions Kanerika’s data engineering team documented the rebuilt logic so it no longer depended on any single analyst’s knowledge. Kanerika migrated the firm’s data estate to Snowflake using the FLIP migration accelerator to automate pipeline conversion and validation. The team rebuilt reconciliation logic on governed, centralized Snowflake data rather than replicating the old spreadsheet process on a new platform.
Results 40% Faster Data Reporting Cycles 3X Quicker Analytics Delivery 60% Reduction in Manual Data Reconciliation
Wrapping Up Snowflake CoCo gives data teams already inside Snowflake a coding agent that understands their actual schemas, permissions, and lineage rather than guessing at them. The Summit 2026 rename brought real product changes alongside it, Cloud Agents, Automations, and a native desktop app, and the pricing model matured from a free preview into standard token billing. It works best as one piece of a broader modernization effort rather than a replacement for one. Teams still reconciling data by hand or running on an undocumented Snowflake estate get more value from fixing that foundation before layering an AI coding agent on top of it.
FAQs What is Snowflake CoCo? Snowflake CoCo is Snowflake’s native AI coding agent, formerly named Cortex Code before its Summit 2026 rebrand. It reads an account’s schemas, RBAC permissions, and lineage before generating SQL, dbt models, Airflow DAGs, or ML pipelines, then plans and executes multi-step tasks rather than only suggesting code snippets. It runs inside Snowsight, a desktop app, or a CLI.
Is Snowflake CoCo the same as Cortex Code? Yes. Snowflake confirms CoCo is simply the new name for Cortex Code, announced at Summit 2026 on June 2. The underlying architecture, capabilities, and pricing model carried over unchanged. Anyone who built workflows around Cortex Code before the rename keeps the same behavior under the new name, with no migration required.
How much does Snowflake CoCo cost? Both Snowsight and CLI usage now bill against token-based AI Credits, after an earlier free preview period for Snowsight ended in early 2026. Costs scale with conversation length and complexity, though repeated context from the second turn onward is served from cache at a discount. CoCo also triggers standard Cloud Services compute for metadata queries.
What is the difference between Snowflake CoCo and Cortex Analyst? CoCo is a coding agent that writes and executes SQL, dbt models, and pipelines from natural language. Cortex Analyst is a separate Snowflake Cortex AI capability built for business users to ask natural-language questions over structured data through a semantic layer. Cortex Agents can orchestrate between the two depending on the request.
Can Snowflake CoCo work inside VS Code or Cursor? Yes. Snowflake ships a dedicated CoCo extension for VS Code, and the CLI integrates into any terminal-capable editor including Cursor. Teams typically use Cursor or VS Code for general editing and run the CoCo CLI in an integrated terminal specifically for Snowflake-aware tasks like schema-grounded SQL generation and dbt work.
Is Snowflake CoCo secure enough for regulated industries? CoCo respects row-level security, masking policies, and existing RBAC, which Snowflake positions as suitable for financial services, healthcare, and insurance workloads. The CLI still carries meaningful risk around credential exposure and shell access, so security teams should containerize it and apply network egress controls rather than relying on governance settings alone.
How does Snowflake CoCo compare to Claude Code and GitHub Copilot? CoCo’s advantage is depth of Snowflake-specific context, schemas, RBAC, and lineage, that neither Claude Code nor GitHub Copilot has by default. Claude Code and Copilot cover general software engineering across any codebase more broadly. Many teams run CoCo alongside one of the other tools rather than choosing a single agent for every task.
Do I need special permissions to use Snowflake CoCo? Yes. An ACCOUNTADMIN must grant SNOWFLAKE.COPILOT_USER to every intended user, plus either SNOWFLAKE.CORTEX_USER or SNOWFLAKE.CORTEX_AGENT_USER depending on workflow complexity. The account also needs cross-region inference enabled. CoCo is unavailable on Government, VPS, and Sovereign Snowflake deployments, so accounts in those categories cannot enable it yet.
