Snowflake CoCo, formerly called Cortex Code, is Snowflake’s AI coding agent for data teams, rebranded at Snowflake Summit 2026 on June 2, 2026. CoCo understands your Snowflake environment natively: schemas, RBAC policies, lineage, and compute costs. It translates natural language into production-ready SQL, dbt models, Airflow DAGs, and ML pipelines, then executes them end-to-end. When paired with a context layer for AI agents, SQL accuracy improves by 3x on complex queries (BirdBench, 145 queries, p < 2e-10, Atlan research 2026).[8] Over 7,100 Snowflake accounts have adopted it, representing more than 50% of the customer base.[3]
| Property | Detail |
|---|---|
| Former name | Cortex Code (launched November 2025) |
| Current name | Snowflake CoCo (rebranded June 2, 2026 at Snowflake Summit) |
| Product type | Data-native AI coding agent |
| Delivery modes | Snowsight, CoCo Desktop (macOS/Windows), CLI, VS Code Extension, Claude Code plugin, Excel Extension, SDK/MCP |
| ADE-Bench pass rate | 72.1% (vs 65.1% for Claude Code and OpenAI Codex)[1] |
| Adoption | 7,100+ accounts; 50%+ of Snowflake customers[3] |
| Model options | Claude Opus 4.7, Claude Sonnet 4.7, GPT 5.4 |
| Pricing | $20/month after 30-day trial; $40 free inference credits for trial[2] |
| New at Summit 2026 | Cloud Agents, Automations, CoCo Desktop, Slack integration, Claude Code plugin, VS Code Extension |
| Security note | Prompt injection CVE patched in CLI v1.0.25 (February 28, 2026); Cortex AI Guardrails added |
Build Your AI Context Stack
Get the blueprint for implementing context graphs across your enterprise. This guide walks through the four-layer architecture — from metadata foundation to agent orchestration — with practical implementation steps for 2026.
Get the Stack GuideWhat is Snowflake CoCo?
Permalink to “What is Snowflake CoCo?”CoCo is Snowflake’s response to a gap that general-purpose AI agent coding tools leave open: they know how to write code, but they do not know your data environment. CoCo fills this by grounding every output in the actual objects, relationships, and governance rules that exist in the connected Snowflake account. It treats the warehouse as context, not just a target for queries.
The key distinction is what “data-native” actually means in practice:
- CoCo reads catalog metadata, column-level lineage, and RBAC policies before generating code
- It routes automatically to the right underlying tool: Cortex Analyst for structured SQL, Cortex Search for semantic object lookup, Semantic Views for business metric definitions
- It references actual objects by their correct names with their correct permissions, rather than best-guess schema approximations
The rebrand: Cortex Code to CoCo
Permalink to “The rebrand: Cortex Code to CoCo”If you find references to “Cortex Code” in documentation, blog posts, or partner materials, including Snowflake Cortex Code Explained, they describe the same product. Snowflake did not change the underlying architecture when it rebranded. It expanded the delivery ecosystem and added new capabilities at Summit 2026.
| Timeline | Event |
|---|---|
| November 2025 | Cortex Code launches; Snowsight and CLI; early SQL and dbt support |
| February 2, 2026 | Cortex Code CLI reaches general availability; dbt and Airflow support added |
| April 2026 | Expanded to “control plane for the agentic enterprise” positioning |
| June 2, 2026 | Rebranded to CoCo at Snowflake Summit 2026 with Cloud Agents, Automations, and ecosystem integrations[3] |
The rename introduced new capabilities alongside the new brand: CoCo Desktop (a native app with persistent project context), Slack integration (coming soon), the Claude Code plugin, VS Code Extension, Excel Extension, Agent Teams for multi-agent orchestration, and Cortex Sense as a runtime context enrichment engine. The Snowflake Summit 2026 announcements are covered in detail in the capabilities section below.
How does Snowflake CoCo work?
Permalink to “How does Snowflake CoCo work?”CoCo operates through a four-step planning loop that distinguishes it from autocomplete-style assistants where a human executes every suggestion.[4]
- Interpret intent: CoCo parses the natural language request, identifies the target output type, and checks the catalog for relevant tables and their RBAC context
- Select tools and build a plan: CoCo selects the appropriate internal tools (Cortex Analyst for SQL logic, Cortex Search for object discovery, lineage metadata for freshness validation) and produces a step-by-step plan with estimated token cost
- Show the plan for review: before any execution begins, the plan is presented to the user, who can edit, reject, or approve each step
- Execute and maintain session context: steps run sequentially; CoCo maintains state across the session without losing context between steps
To make this concrete: if you ask CoCo to “build a dbt model for daily customer churn,” it checks which tables in your Snowflake account contain customer and event data, verifies which ones you have read access to under your RBAC policy, confirms their upstream freshness signals, drafts the dbt model logic, shows you the full plan before touching anything, and then (after your approval) generates, tests, and stages the model.
This is what practitioners mean by “human-on-the-loop”: the agent does the work, the human supervises. GitHub Copilot is “human-in-the-loop”: the agent suggests, the human executes. See how to build an AI agent for the full framework that makes these agents production-ready.[5] These are structurally different product categories, not versions of the same approach.
Multi-agent execution with Agent Teams
Permalink to “Multi-agent execution with Agent Teams”Summit 2026 added an Agent Teams layer that shifts CoCo from sequential planning to concurrent orchestration. A lead agent assigns sub-agents to parallel roles (research, coding, testing); results are synthesized before being returned. For complex data engineering tasks, this meaningfully reduces total execution time.
Delivery modes
Permalink to “Delivery modes”CoCo is available across multiple surfaces, a deliberate expansion at Summit 2026:
| Mode | Best for |
|---|---|
| Snowsight (web) | In-warehouse workflows, notebooks, ad-hoc SQL, analyst users |
| CoCo Desktop (macOS/Windows) | Full agentic development; persistent project context across sessions |
| CLI | Terminal-native development, CI/CD pipelines, local automation |
| VS Code Extension | IDE-native integration for developers |
| Claude Code plugin | Cross-tool development workflows |
| SDK and MCP server | Programmatic embedding into custom workflows |
Internal tools CoCo routes to
Permalink to “Internal tools CoCo routes to”CoCo is not a single model. It routes task execution to Snowflake’s internal tool stack automatically:
- Cortex Analyst: structured query engine for reliable, schema-grounded SQL generation
- Cortex Search: semantic search across Snowflake objects
- Snowflake Horizon Context and semantic views: business metric definitions within Snowflake (the Snowflake-native answer to “what does ‘revenue’ mean in this account?”)
- Catalog and lineage metadata: actual table ownership, column origin, and freshness signals, including Snowflake data lineage signals
- RBAC context: ensures generated code references only the tables the user is authorized to access
| Dimension | Traditional SQL assistant | Snowflake CoCo |
|---|---|---|
| Context source | General code syntax; no warehouse awareness | Snowflake schemas, RBAC policies, lineage, semantic models, cost patterns |
| Execution model | Suggests; human executes | Plans, reviews, executes end-to-end |
| Environment fit | Syntactically correct; may not match actual schema | Grounded in actual objects with correct permissions |
| Tool routing | None | Auto-routes to Cortex Analyst, Cortex Search, Semantic Views |
| Governance | None native | Runs under existing Snowflake RBAC with full audit trail |
| Session state | Stateless across steps | Maintains session context throughout multi-step execution |
What can Snowflake CoCo do? Key capabilities
Permalink to “What can Snowflake CoCo do? Key capabilities”CoCo generates SQL, Python, dbt models, Airflow DAGs, ML pipelines, and AI agent code from natural language. According to Snowflake’s ADE-Bench results (dbt Labs, June 2026), CoCo achieves a 72.1% pass rate on real-world analytics engineering tasks, compared to 65.1% for both Claude Code and OpenAI Codex, while using 51% fewer tokens and running 8% faster on Opus 4.7.[1]
Data engineering and pipeline building
Permalink to “Data engineering and pipeline building”Core use cases, each grounded in Snowflake-native context:
- Natural language to SQL: complex JOINs, window functions, and CTEs built from actual catalog context, not best-guess schema assumptions
- Natural language to dbt models: CoCo understands dbt project structure, generates models, runs tests, and stages deployment
- Natural language to Airflow DAGs: DAGs that reference actual Snowflake task dependencies, not generic task templates
- ML pipeline scaffolding: feature extraction, model training on Snowflake data, and inference deployment from natural language descriptions
- Data diffing: schema version comparison, drift detection, and migration script generation
- Agent building: generate and deploy Cortex Agents using Snowflake’s open
agents.mdframework — the agent harness pattern makes these agents production-ready
CoCo Automations: recurring event-driven workflows (new at Summit 2026)
Permalink to “CoCo Automations: recurring event-driven workflows (new at Summit 2026)”Automations are the capability that moves CoCo from a productivity tool to a production runtime agent, and they are underexplained in most current coverage.
A CoCo Automation is a recurring, event-driven workflow defined once and executed continuously. You define the trigger (a schedule, a data arrival event, a threshold alert, or a pipeline failure signal), and CoCo builds and deploys the corresponding workflow. Automations run under Snowflake RBAC as the user’s identity, not a generic service account, and produce a full audit trail.
Practical use cases:
- Daily pipeline health checks: CoCo runs your defined validation logic every morning and surfaces anomalies without manual intervention
- Weekly model drift detection: scheduled model quality checks against production data, with alerts when drift crosses a threshold
- Real-time anomaly alerts: event-triggered investigation workflows that run the moment a sensor or quality rule fires
- Automated incident investigation: when a pipeline failure event occurs, CoCo runs a root-cause investigation workflow and returns a structured report
The governance model matters: because Automations run as the user’s identity, they cannot escalate privileges beyond what the defining user was authorized to access at creation time.
Cloud Agents: serverless execution (new at Summit 2026)
Permalink to “Cloud Agents: serverless execution (new at Summit 2026)”CoCo Cloud Agents run in managed cloud environments within Snowsight: isolated containers that support shell commands, Python scripts, package installation, file operations, dbt builds and tests, and web searching. The key architectural implication is that long-running jobs complete in the cloud and report back when finished. A developer’s machine does not need to stay active for a multi-hour dbt build or a large-scale ML training run.
Ecosystem integrations (new at Summit 2026)
Permalink to “Ecosystem integrations (new at Summit 2026)”The Summit 2026 release significantly expanded where CoCo operates:
- Slack integration (CoCo Slackbot, coming soon): launch workflows and check task status from Slack channels
- Claude Code plugin: access CoCo from Anthropic’s Claude Code CLI for cross-tool development
- Microsoft Excel Extension: spreadsheet-based access for analyst-persona users
- SAP skill: dedicated connector to SAP data within Snowflake
- Skill Catalog: share and reuse proven CoCo workflows across teams
Snowflake CoCo vs GitHub Copilot: what’s different for data teams?
Permalink to “Snowflake CoCo vs GitHub Copilot: what’s different for data teams?”GitHub Copilot and Snowflake CoCo are not competing in the same product category. Copilot is a general-purpose code autocomplete tool for software developers; CoCo is a data-native execution agent for data engineers and analysts. The core difference: Copilot has no knowledge of your data environment. CoCo’s entire value proposition is depth of knowledge of your Snowflake environment.[5]
| Dimension | Snowflake CoCo | GitHub Copilot | dbt |
|---|---|---|---|
| Product category | Data-native AI execution agent | General-purpose code autocomplete | SQL transformation framework |
| Execution model | Plans and executes end-to-end (human-on-the-loop) | Suggests; human executes (human-in-the-loop) | Developer writes; dbt compiles and runs |
| Warehouse context | Snowflake schemas, RBAC, lineage, semantic models | None natively (MCP workaround available) | dbt project structure; no RBAC awareness |
| Persona served | Data engineers, analysts, ML engineers, data-savvy business users | Software developers | Data engineers |
| Governance | Native Snowflake RBAC with full audit trail | External access tokens; no native governance | Testing and version control; no RBAC |
| Relationship to CoCo | – | Can be enhanced with Snowflake MCP server | CoCo generates and deploys dbt models; complementary |
| Output fidelity | Production-ready code grounded in actual environment | Syntactically correct; may not fit actual schema | Governed transformation logic with testing |
dbt and CoCo are increasingly complementary rather than competing. CoCo can generate, test, and deploy dbt models through natural language, understanding both the Snowflake environment and the dbt project structure simultaneously. Before CoCo, developers had to generate Snowflake-aware code in one environment and manually transpose it to dbt. CoCo eliminates that gap.
CoCo’s Snowflake-native advantage comes with a structural limit: its context ends where Snowflake’s context ends. Business definitions that span tools, cross-system lineage, and governance signals from outside Snowflake are not natively available — a gap a context layer bridges. The next section addresses that gap.
Security consideration: what data teams should know
Permalink to “Security consideration: what data teams should know”In March 2026, PromptArmor disclosed a prompt injection vulnerability in the Cortex Code CLI: a malicious README could cause CoCo to execute arbitrary commands under the user’s active Snowflake credentials.[6] Snowflake had already patched the vulnerability in CLI v1.0.25 on February 28, 2026, before the public disclosure on March 16, 2026. After patching, Snowflake added Cortex AI Guardrails, layered prompt injection protections, sandboxing improvements, and prompt/response logging with query tagging and admin cost controls.
CoCo is production-safe as patched. The incident surfaced a structural principle that applies to all AI coding agents: agentic execution at scale requires controlled context delivery, not just controlled code.
Any agent that reads external content (READMEs, documentation, data files) and has execute access to production systems carries prompt injection exposure if that content is not validated before reaching the agent. The richer an agent’s context, the more carefully that context must be curated. This is the argument for Context Engineering Studio: validate what the agent sees before it ships to production. Not just the code it writes. The context it reads — because AI agent hallucination risk is highest when agents operate on uncurated context.
How Atlan’s context layer makes CoCo more accurate
Permalink to “How Atlan’s context layer makes CoCo more accurate”CoCo understands Snowflake schemas deeply. It does not understand the business meaning behind them. According to Atlan research (2026), 49% of organizations report that AI underperformance stems from missing context, not model limitations. CoCo is not the exception.
What CoCo lacks natively:
- The business glossary definition of “customer” as agreed by Finance, GTM, and Product, which may differ meaningfully from what a column name suggests
- Column-level lineage from Fivetran ingestion to Tableau consumption: what happens to data before it reaches Snowflake, and where it goes after
- Governance signals and ownership metadata from dbt metrics, Looker explores, or BI layers
- Data quality signals from upstream systems
- The context graph: the living map of how business concepts connect across tools and teams
Atlan is the context layer for Snowflake: a complementary layer on top of Snowflake’s data plane that fills this gap. The context layer for Snowflake Cortex guide covers what’s needed for production-grade CoCo accuracy.
How the context delivery chain works
Permalink to “How the context delivery chain works”Atlan’s four products work in sequence to deliver enterprise context to CoCo, functioning as a context layer as AI memory that persists and compounds knowledge:
- Enterprise Data Graph (100+ connectors): pulls metadata from every system in the stack, including Snowflake, dbt, Fivetran, Airflow, Looker, Tableau, Jira, and Confluence. Reverse-engineers column-level lineage from SQL. Builds cross-system ownership and quality signals — the foundation of an enterprise data graph.
- context agents: AI teammates that auto-generate descriptions, metrics, glossary terms, and ontology from the Enterprise Data Graph. 690K+ descriptions generated; 87% rated on par or better than human writing across 50+ enterprise customers — see context agents for methodology.
- Context Engineering Studio: bootstraps, tests, and versions the context CoCo will receive. CI-integrated eval suites validate context before it ships to production. Context repos function as Git repos for context: versioned, auditable, and rollback-capable.
- Atlan MCP server: delivers validated context from the Context Lakehouse (Atlan’s Iceberg-native context store, engineered natively for AI) to CoCo at runtime via Model Context Protocol, alongside Snowflake’s native schema context.
Snowflake has its own answer to context quality: Cortex Sense, announced at Summit 2026 and entering private preview, builds runtime signals from Snowflake activity and personalizes context for CoCo based on user behavior. Cortex Sense reaches within the warehouse.[9] Atlan’s Enterprise Data Graph reaches across every system in the stack. Both layers are complementary: Cortex Sense provides warehouse-native signal enrichment; Atlan provides the cross-system business context layer. Learn more in Snowflake Horizon 101 and how Snowflake Horizon Catalog and Atlan work together.
| Context type | CoCo native | Atlan Enterprise Data Graph adds |
|---|---|---|
| Schema / table / column definitions | Yes (Snowflake objects) | Cross-system; upstream sources including Fivetran, Airflow |
| RBAC / access policies | Yes (Snowflake RBAC) | Cross-system governance propagation; two-way Horizon tag sync |
| Data lineage | Snowflake-native | Column-level, cross-system (Airflow, dbt, Tableau, etc.) |
| Business glossary | Not available | Yes: curated, team-approved definitions across all tools |
| Semantic definitions (metrics) | Snowflake Semantic Views | dbt metrics, Looker explores, BI definitions mapped to Snowflake semantics |
| Data quality signals | Limited | Cross-system quality metrics from upstream systems |
| Ownership / stewardship | Not available | Yes: per asset, with contact info |
| Cortex Sense enrichment | Snowflake activity signals | Atlan adds context from outside Snowflake’s perimeter |
The 3x accuracy result
Permalink to “The 3x accuracy result”In a controlled BirdBench study, Atlan tested 145 queries comparing metadata-enriched context. Context engineering principles systematize this enrichment: (Atlan business definitions + Snowflake schema) against schema-only context. According to Atlan research (2026), SQL generation accuracy improved 3x on complex or ambiguous queries (p < 2e-10).[8] The gap is largest where query intent depends on knowing what a business term means, not just what a column is named.
The formula: CoCo’s technical context (schemas, RBAC, Snowflake-native lineage) + Atlan’s enterprise context (cross-system lineage, business glossary, certified metrics, governance) = an agent that codes with full business context, not just technical context.
Atlan is Snowflake’s 2025 Data Governance Partner of the Year and a launch partner for Snowflake Intelligence, bringing AI agent governance to the integration. Explore how Snowflake Intelligence and Atlan work together and the MCP server for Snowflake that delivers context at runtime.
Inside Atlan AI Labs & The 5x Accuracy Factor
Learn how context engineering drove 5x AI accuracy in real customer systems. Explore real experiments, quantifiable results, and a repeatable playbook for closing the gap between AI demos and production-ready systems.
Download E-BookReal stories from real customers: context in production
Permalink to “Real stories from real customers: context in production”"We're excited to build the future of AI governance with Atlan. All of the work that we did to get to a shared language at Workday can be leveraged by AI via Atlan's MCP server…as part of Atlan's AI Labs, we're co-building the semantic layer that AI needs with new constructs, like context products."
— Joe DosSantos, VP of Enterprise Data & Analytics, Workday
"Atlan is much more than a catalog of catalogs. It's more of a context operating system…Atlan enabled us to easily activate metadata for everything from discovery in the marketplace to AI governance to data quality to an MCP server delivering context to AI models."
— Sridher Arumugham, Chief Data & Analytics Officer, DigiKey
Why CoCo and a context layer work better together
Permalink to “Why CoCo and a context layer work better together”CoCo is an excellent AI agent within Snowflake’s walls. It outperforms general-purpose coding assistants on data engineering tasks, executes end-to-end with governance built in, and (as of Summit 2026) runs in serverless environments and triggers on schedules without any developer session staying open.
The limit is the same as the strength: CoCo’s context ends where Snowflake’s context ends — this is why a context layer for Snowflake Cortex matters for cross-system accuracy. For a data team that lives entirely in Snowflake, that limit may rarely surface. For most enterprises (where data moves through Fivetran, Airflow, dbt, Looker, Tableau, and other systems before anyone queries it in Snowflake) the missing context is exactly where the hardest queries live.
Adding Atlan’s Enterprise Data Graph as the context plane on top of Snowflake’s data plane closes that gap. CoCo still writes the code. Atlan ensures the agent understands what the code should mean.
FAQs about Snowflake CoCo
Permalink to “FAQs about Snowflake CoCo”1. Is Snowflake CoCo the same as Cortex Code?
Yes. Snowflake rebranded Cortex Code to CoCo at Snowflake Summit 2026 on June 2, 2026. The underlying technology is the same product. Documentation may still reference “Cortex Code” in older pages; both names refer to the same AI coding agent.
2. How do I access Snowflake CoCo?
CoCo is available in Snowsight (Snowflake’s web UI), as a native desktop application (CoCo Desktop for macOS and Windows), via CLI, via the VS Code Extension, and via the Claude Code plugin. A 30-day free trial with $40 in inference credits is available; the subscription is $20/month after trial.
3. What is CoCo Desktop?
CoCo Desktop is a native macOS and Windows application for full agentic development outside the browser. It maintains persistent project context across sessions, meaning CoCo retains awareness of your schema decisions, table choices, and prior work without needing to re-establish context on each use.
4. What are CoCo Automations?
CoCo Automations are recurring, event-driven workflows that run on a schedule or in response to a trigger (data arrival, threshold alert, pipeline failure). Examples include daily pipeline health checks, weekly model drift detection, and automated anomaly investigation. Automations are governed by Snowflake RBAC with full audit trails; they run as the user’s identity, not a generic service account.
5. What are CoCo Cloud Agents?
CoCo Cloud Agents are managed cloud execution environments within Snowsight. They run tasks (Python scripts, dbt builds, shell commands, file operations) in isolated containers without requiring a local session to stay open. Long-running jobs complete in the cloud and report results back.
6. What is Cortex Sense, and how does it relate to CoCo?
Cortex Sense is Snowflake’s runtime context enrichment layer, announced at Summit 2026 and entering private preview. It builds signals from data and activity already in Snowflake and automatically personalizes context for CoCo based on user role and behavior patterns. Cortex Sense enhances CoCo within the Snowflake perimeter; it does not reach cross-system business definitions, external lineage, or governance signals from outside Snowflake.
7. Is Snowflake CoCo safe to use in production?
Yes, with standard precautions. A prompt injection vulnerability in the CLI was patched in CLI v1.0.25 on February 28, 2026, before public disclosure on March 16, 2026. Snowflake added Cortex AI Guardrails, prompt injection protections, and improved sandboxing. CoCo operates under existing Snowflake RBAC and cannot access tables the authenticated user cannot access. Prompt and response logging and query tagging are available for audit.
8. What models does Snowflake CoCo use?
CoCo is model-agnostic by design. As of Summit 2026, the available model options are Claude Opus 4.7, Claude Sonnet 4.7, and GPT 5.4. CoCo also supports intelligent cost/accuracy routing across model options, so teams can balance output quality against inference cost per task.
Sources
Permalink to “Sources”- Snowflake CoCo: AI Coding Agent for the Modern Data Stack, Snowflake Blog
- Snowflake CoCo Product Page, Snowflake
- Snowflake CoCo Redefines Enterprise AI Development, Snowflake Press Release
- Cortex Code Documentation, Snowflake Docs
- Snowflake Cortex Code Is Not a Copilot, Medium
- Snowflake Cortex AI Escapes Sandbox and Executes Malware, PromptArmor
- What Snowflake Cortex Needs From a Context Layer, Atlan
- Making Talk to Data 3x More Accurate: Snowflake Intelligence and Atlan, Atlan
- Under the Hood at Snowflake Summit 2026: The Engineering Leaps Powering the Agentic Enterprise, iTWire
- Snowflake Summit 2026: Context, custom model training, Iceberg V3, Constellation Research
