06-reference/research

snowflake eai privileges and coding agent sandbox egress

2026-11-10·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
spcsexternal-access-integrationcortex-codesnowflakephdata

The ACCOUNTADMIN ask for SPCS egress can be one-time, and a no-secret EAI is documented. The Coding Agent sandbox has no documented EAI hook, so treat its egress as fixed; CoCo automations get GitHub and Snowflake-registered MCP servers, nothing else

The question

Part A (verbatim): "For brigade-in-SPCS egress: which privileges does a deploy role need to create the network rule and secret itself versus ACCOUNTADMIN pre-creating them, and does Snowflake document an external access integration with NO ALLOWED_AUTHENTICATION_SECRETS?"

Part B (verbatim): "Can an external access integration be attached to the GA Cortex Agents Coding Agent sandbox, or is its egress fixed by Snowflake?" Per the merge instruction, Part B also answers the egress half only of the adjacent item "can the same sandbox or CoCo (Cortex Code) automations run the brigade's unattended service loop". The subagent/hook/plugin loading and compute half stays open.

Both questions are follow-ups from [[2026-09-26-spcs-brigade-external-access-anthropic-reachability]]. Part A decides whether the client-admin ask is one-time or recurring. Part B decides whether the Coding Agent sandbox survives as the P3 upgrade path named in [[2026-09-13-cortex-code-client-deployable-surface-p3-p4]].

What we already know (from the vault)

What the web says

All pages below were fetched on 2026-11-10 directly from docs.snowflake.com, either the HTML or the .md rendition. Reference pages carry no visible page date. Release-note dates are given where they exist. Evidence grades: documented means stated in so many words. Ambiguous means the docs speak to it but conflict or are scoped elsewhere. Undocumented means silent.

Part A: privileges and the no-secret EAI

Part B: Coding Agent sandbox and CoCo automations egress

Convergences and contradictions

Synthesis for RDCO

Part A: the ask can be one-time, but only through a deliberate ownership split, and that split carries a security trade-off the client admin should choose knowingly. A deploy role can create its own network rules and secrets without any recurring ACCOUNTADMIN involvement. Both are schema privileges, and the schema owner holds CREATE NETWORK RULE by default. So if the deploy role owns a dedicated egress schema, it needs no extra grant for the rule, and CREATE SECRET on that schema is one ordinary grant. The only account-level object is the EAI. That gives two designs:

Recommendation for the client conversation: offer design (1) by default and design (2) as the low-friction option. Either way, name the host list up front (git remote, any MCP hosts, package mirrors if the image is not prebuilt). The model path needs no host if inference goes through Cortex.

On secrets: a no-secret EAI is documented and is the right default for the brigade. If the brigade runs on the CoCo CLI inside the container, inference goes through Cortex and there is no Anthropic key at all. The container gets its Snowflake session token from SPCS, so the EAI exists purely to allowlist boring hosts, and ALLOWED_AUTHENTICATION_SECRETS can be omitted. Where a credential truly is needed (a git PAT, for example), the sanctioned handoff is a Snowflake SECRET referenced in containers.secrets. The spec's envVarName + secretKeyRef: secret_string option exposes it as an environment variable; directoryPath writes it as a file. This is the client-side equivalent of our 1Password wrappers: no .env file in the image, rotation in Snowflake, and READ gated by RBAC. However, it does put the plaintext value in the container's environment, unlike the GitHub proxy injection Snowflake built for CoCo sandboxes. Whether that secret also has to be listed on the EAI is a 15-minute live test, not a design blocker.

Part B: treat the Coding Agent sandbox's egress as fixed by Snowflake. No EAI can be attached according to anything documented. Its only egress levers are package registries (via the Artifact Repository on the sibling tool), web_search, and Snowflake itself. It has no --github equivalent. So it cannot host any brigade station that pushes to git or talks to a non-Snowflake MCP server. That disqualifies it as the P3 upgrade path for the full brigade, as the backlog notes predicted. One escape hatch is documented piece by piece but not end to end: wrap each non-Snowflake call in an EAI-bearing stored procedure, expose it as a GENERIC tool on a Snowflake-managed MCP server, and let the sandboxed agent call that tool. The egress then happens in a warehouse under a normal EAI, so the client's admin sees the same object set as design (1). This works for a station that needs one or two external calls. It is a poor fit for a station whose job is git.

On the merged item: CoCo automations are a better egress fit than the Coding Agent, but only for GitHub-hosted repos and Snowflake-registered MCP servers. They got GitHub egress with proxy-injected tokens (--github), pre-run and post-run hooks that can git clone and git push, and --mcp for MCP server objects. They went GA on 2026-11-02, in commercial regions only. Clients on GitLab, Bitbucket or Azure DevOps have no documented path. A brigade MCP server hosted outside Snowflake must be fronted by a Snowflake MCP server object. Two client-side settings can switch all of this off. COCO_CLOUD_AGENTS_NON_SNOWFLAKE_EGRESS_DISABLED, if it applies to automations (ambiguous), removes even GitHub and package registries. Revoking EXECUTE AGENT TASK from PUBLIC removes automations entirely. Both belong on the P3 discovery checklist next to CORTEX_ENABLED_CROSS_REGION. Net effect on the matrix: P3 stays "CoCo CLI in SPCS" for the full brigade. Automations become a credible host for the egress needs of GitHub-only, MCP-in-Snowflake service loops, pending the open loading-and-compute half. The Coding Agent drops to "embed a single station via API", not "host the brigade".

Why this is in the vault

It settles the egress and grant cells of the P3 row in [[2026-07-07-hex-deployment-matrix-v2]]. It also gives the client-admin script for brigade-in-SPCS: two schema grants, one admin-created EAI, a named ownership split, and no Anthropic secret. Finally, it removes the Coding Agent sandbox as the P3 upgrade path that [[2026-09-13-cortex-code-client-deployable-surface-p3-p4]] told us to watch, and puts CoCo automations in that slot for GitHub-hosted clients.

Open follow-ups

Related

Sources