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)
- The EAI object set is a
NETWORK RULE(egress, host-port), an optionalSECRET, and theEXTERNAL ACCESS INTEGRATIONthat binds them.CREATE INTEGRATION ON ACCOUNTis the account-level grant, and ACCOUNTADMIN is not required for the public-internet case. The schema-level grants for the network rule were left unpinned. [[2026-09-26-spcs-brigade-external-access-anthropic-reachability]] - That brief also said "the docs show the secret-bearing pattern only". This brief finds that statement was wrong; see the corrections note in the Convergences section.
- Every CoCo surface routes model calls through Cortex, so a CoCo-based brigade needs no
api.anthropic.comegress. Residual egress is git, non-Snowflake MCP servers and package installs. [[2026-09-13-cortex-code-client-deployable-surface-p3-p4]] - On 2026-09-13 the vault recorded the Coding Agent (
code_toolset_all, GA 2026-08-26) as a "Snowflake-managed sandbox on the CoCo runtime" with unconfirmed extensibility. It also listed CoCo automations as Preview and Cloud Agents as private preview. Both statuses are now out of date (see below). [[2026-09-13-cortex-code-client-deployable-surface-p3-p4]] - Injecting a credential through an egress proxy is a known pattern: the proxy adds the token on the way out, so code inside the sandbox never holds it. [[2026-05-23-harrison-chase-langsmith-auth-proxy]]. Snowflake now documents this pattern, but only for GitHub in CoCo sandboxes (Part B).
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
CREATE NETWORK RULEis a schema privilege that the schema owner holds by default (documented). CREATE NETWORK RULE, access control table: "CREATE NETWORK RULE | Schema | Only the ACCOUNTADMIN and SECURITYADMIN roles, along with the schema owner, have this privilege by default. It can be granted to additional roles as needed." The privileges reference lists it under schema privileges: "Enables creating a new network rule in a schema." Changing an existing rule takesOWNERSHIPon the rule. This closes the gap left open in the parent brief.CREATE SECRETis also schema-level (documented). CREATE SECRET: "CREATE SECRET | Schema", plus "USAGE | Database or schema". Both pages add: "Operating on an object in a schema requires at least one privilege on the parent database and at least one privilege on the parent schema."Only the EAI itself is account-level, and the CREATE EAI page lists a narrower grant than the parent brief did (documented, but scope is ambiguous). The CREATE EXTERNAL ACCESS INTEGRATION access table lists four rows:
- "CREATE INTEGRATION | Account | Only the ACCOUNTADMIN role has this privilege by default."
- "USAGE | Secret | Required for all secrets referenced by the integration."
- "USAGE | Schema | Required for all schemas containing any secrets referenced by the integration."
- "CREATE EXTERNAL ACCESS INTEGRATION | Account | Grants the ability to create external access integrations. This privilege does not grant the ability to create other types of integrations."
The privileges reference describes that same granular privilege differently: "Grants a Snowflake Native App the ability to create an external access integration." Its
CREATE INTEGRATIONrow says "To let a role create only one type of integration, grant the corresponding granular privilege instead." Whether a normal account role can holdCREATE EXTERNAL ACCESS INTEGRATIONis ambiguous across the two pages. The table does not list any privilege on the referenced network rules (undocumented).SPCS docs assume the admin creates the EAI, and they forbid ACCOUNTADMIN from creating the service (documented). SPCS service networking: "Typically, you want an account administrator to create EAIs ... Account administrators can then grant EAI usage to specific roles that developers use to run services." The same page says: "The owner role that is creating the service needs the USAGE privilege on the EAI and READ privilege on the secrets referenced. Note that you cannot use the ACCOUNTADMIN role to create a service."
An EAI with no secrets is documented, both in the syntax and in a worked SPCS example (documented). The CREATE EAI syntax has
[ ALLOWED_AUTHENTICATION_SECRETS = { ( <secret_name> ... ) | all | none } ]under Optional parameters, with "none to allow no secrets." The SPCS networking page gives the full exampleCREATE EXTERNAL ACCESS INTEGRATION google_apis_access_integration ALLOWED_NETWORK_RULES = (translate_network_rule, google_network_rule) ENABLED = true;, thenGRANT USAGEandCREATE SERVICE ... EXTERNAL_ACCESS_INTEGRATIONS = (...), with no secret anywhere. The parameter's own description is scoped to "UDF or procedure handler code".The sanctioned way to hand an API key to an SPCS container is
containers.secrets(documented). Service specification reference: "Use the containers.secrets field in your service specification to provide Snowflake-managed credentials to your application containers." Fields aresnowflakeSecretplus eitherenvVarName(withsecretKeyRef: username | password | secret_string) ordirectoryPath, "to pass the secret as environment variables or ... to write the secrets to local container files." The service owner role "will need the READ privilege on the secrets referenced." Whether aGENERIC_STRINGsecret passed this way must also be listed in the EAI'sALLOWED_AUTHENTICATION_SECRETSis ambiguous. The spec reference requires only READ, and the parent brief quoted the must-be-allowed rule for OAuth secrets only.
Part B: Coding Agent sandbox and CoCo automations egress
- The Coding Agent page documents no EAI attachment point (undocumented, which in practice means fixed). Coding Agent documents only these
tool_resources.code_toolset_allkeys:permission_policy,workspace_mounts,stage_mounts(added 2026-10-28), anddisabled_skills. It does not use the words network, egress, internet or external access. Its only pointer is: "For thecode_executiontool, default libraries, adding libraries through the Artifact Repository, and external network access, see [Cortex Agent code execution tool]." - The page that pointer leads to has no EAI either (documented-by-omission). Code execution tool contains the word "external" zero times. It says: "By default, the code execution tool environment is isolated and can only access data passed into the session, plus the workspace stage mounted into the sandbox." Its only egress lever is
artifact_repositories: [SNOWFLAKE.SNOWPARK.PYPI_SHARED_REPOSITORY]plus theSNOWFLAKE.PYPI_REPOSITORY_USERrole. - Two GA pages describing the same runtime family contradict each other on EAIs (ambiguous).
- Cloud Agents (Snowsight, now GA) says: "HTTPS egress is limited to a fixed allowlist of package managers and related build tooling ... Hosts outside that allowlist are not reachable from the sandbox. External access integrations are not currently supported for Cloud Agents." The allowlist is read-only: "GET and HEAD only ... publish and other write operations are blocked."
- CoCo CLI cloud sandbox (GA;
cortex --cloud) says: "Access to other hosts requires external access integrations that your Snowflake administrator configures. These use the same primitives as UDFs and stored procedures." It gives no flag, parameter or SQL that binds an EAI to the sandbox.
- GitHub is the one non-registry host with a documented path, and it uses proxy injection rather than a credential in the container (documented). From the cloud-sandbox page: "Snowflake fetches the token server-side using your session, and injects it only on outbound requests to GitHub. The token never enters the container's environment in plaintext." The
--github <secret_fqn>flag exists forcortex --cloudand for automations. It does not exist in the Coding Agent'stool_resources. - An account kill switch exists, and it is ACCOUNTADMIN-only (documented). Setting
COCO_CLOUD_AGENTS_NON_SNOWFLAKE_EGRESS_DISABLED = TRUEmeans "the container cannot reach hosts outside Snowflake, including package registries and GitHub" (cloud-sandbox page). The parameters reference scopes it to "the Cortex Code Cloud Agents sandbox". Whether it also covers the Coding Agent and automation sandboxes is ambiguous. - CoCo automations went GA on 2026-11-02 as "recurring, unattended CoCo runs that execute in a Snowflake-managed sandbox". Their egress surface is GitHub plus MCP servers that are registered as Snowflake objects (documented). CoCo automations lists the options
--github <secret_fqn>("authenticated GitHub egress") and--mcp <database.schema.name>("Attach a Snowflake-managed or customer-provided MCP server"). It also says: "Only files mounted from Snowflake and MCP servers explicitly attached to the automation are available." There is no EAI option. "Customer-provided" is not defined, but the argument is a Snowflake object name. The page documents a hook-basedgit clone/git pushagainst GitHub, and it warns that "a broken hook can go unnoticed". - One documented indirection does reach arbitrary hosts (documented parts, inferred composition). CREATE MCP SERVER supports
GENERICtools backed by UDFs or stored procedures that run on awarehouse. UDFs and procedures can carryEXTERNAL_ACCESS_INTEGRATIONS(shown in the CREATE EAI example). ACODING_AGENTMCP tool (2026-10-23) "appendscode_toolset_allafter any tools inconfig". Putting those together: a non-Snowflake call can run outside the sandbox, as an EAI-bearing procedure tool. No page shows this chain end to end.
Convergences and contradictions
- Correction to the parent brief. [[2026-09-26-spcs-brigade-external-access-anthropic-reachability]] said "the docs show the secret-bearing pattern only". The SPCS networking page, which that brief itself cited through a search summary, contains a no-secret EAI example. The CREATE EAI syntax also accepts
none. The parent's instinct that a no-secret EAI should work was right. Only its claim about the docs was wrong. - The vault's grant list was close but incomplete. The parent brief named
CREATE INTEGRATIONas the account-level ask. The current CREATE EAI page addsUSAGEon every referenced secret and its schema, which the EAI creator needs. It also offers a granularCREATE EXTERNAL ACCESS INTEGRATIONprivilege that would be the minimal-blast-radius ask, if it can be granted to a non-app role. - The main contradiction is between two Snowflake GA pages. Cloud Agents says EAIs are "not currently supported". The CLI cloud sandbox says other hosts "require" EAIs but shows no way to bind one. The Coding Agent page says nothing on the subject. Three surfaces sit on "the same runtime that powers CoCo", and they give three different answers. The conservative reading is that today no customer-attachable EAI exists for any managed CoCo sandbox. The CLI sentence reads as a roadmap or generic statement, because it has no syntax behind it.
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:
- (1) The admin owns the EAI and the rule. The admin creates the rule and the EAI once, then grants
USAGEon the EAI to the deploy role. Every new host (a new git remote, a new MCP server) is another ACCOUNTADMIN ticket. That is recurring, but the allowlist stays under the client's control. - (2) The admin owns the EAI, and the deploy role owns the rule. The admin creates the EAI once, pointing at a network rule that the deploy role owns. After that, the deploy role adds hosts with
ALTER NETWORK RULE(anOWNERSHIPprivilege) and never asks again. The cost is that the deploy role can widen egress on its own, up to0.0.0.0:443. - (3) The deploy role creates the EAI itself. This needs
CREATE INTEGRATION, which also lets the deploy role create storage, security and API integrations. That is a worse blast radius than (2), and phData should not ask for it. If the granularCREATE EXTERNAL ACCESS INTEGRATIONturns out to be grantable to a normal role, it becomes the clean version of (3).
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
- Live-account tests, not research questions. These settle the remaining ambiguities on a phData sandbox account:
- (a)
GRANT CREATE EXTERNAL ACCESS INTEGRATION ON ACCOUNT TO ROLE deploy_role: does it succeed for a non-app role? - (b) ACCOUNTADMIN creates an EAI that references a rule owned by the deploy role; the deploy role then runs
ALTER NETWORK RULEto add a host. Does a running service reach the new host without changing the EAI or restarting the service? - (c) An EAI with no
ALLOWED_AUTHENTICATION_SECRETS, plus aGENERIC_STRINGsecret passed throughcontainers.secrets.envVarName. DoesCREATE SERVICEsucceed? - (d) A Coding Agent request with
bash: curl -sI https://example.comandhttps://api.github.com, and with an unknowntool_resources.code_toolset_all.external_access_integrationskey. Is the request rejected, or is the key silently ignored?
- (a)
- Does
COCO_CLOUD_AGENTS_NON_SNOWFLAKE_EGRESS_DISABLEDgovern the Coding Agent and automation sandboxes, or only Cloud Agents andcortex --cloud? The parameter reference and the cloud-sandbox page scope it differently. - Has Snowflake published (or will it publish) syntax that binds an EAI to a CoCo cloud sandbox? The CLI cloud-sandbox page promises it, and the Cloud Agents page denies it. Re-check the release notes in about 30 days.
- What does "customer-provided MCP server" mean for
--mcpin automations? Can a Snowflake MCP server object proxy a remote, non-Snowflake MCP endpoint directly, or only throughGENERICUDF and procedure tools?
Related
- [[2026-09-26-spcs-brigade-external-access-anthropic-reachability]]
- [[2026-09-13-cortex-code-client-deployable-surface-p3-p4]]
- [[2026-07-07-hex-deployment-matrix-v2]]
- [[2026-07-12-spcs-app-runtime-enablement-partner-managed]]
- [[2026-07-05-brigade-skills-snowflake-cortex-code]]
- [[2026-05-23-harrison-chase-langsmith-auth-proxy]]
Sources
- Vault: [[2026-09-26-spcs-brigade-external-access-anthropic-reachability]] (parent brief; source of both Part A follow-ups and the Part B question)
- Vault: [[2026-09-13-cortex-code-client-deployable-surface-p3-p4]] (P3 upgrade-path framing; prior Coding Agent, automations and Cloud Agents status)
- Vault: [[2026-07-07-hex-deployment-matrix-v2]] (P3 row this brief updates)
- Vault: [[2026-07-12-spcs-app-runtime-enablement-partner-managed]] (original account-level grant list)
- Vault: [[2026-07-05-brigade-skills-snowflake-cortex-code]] (C1 egress framing)
- Vault: [[2026-05-23-harrison-chase-langsmith-auth-proxy]] (auth-proxy credential-injection pattern)
- Research caps: this brief deliberately went past the template's 3-fetch-per-question cap. It fetched 16 primary docs.snowflake.com pages via curl, because the task required exact quotes and the decisive pages (cloud sandbox, Cloud Agents) were only reachable by following the docs index. It used 0 WebSearch queries and 1 QMD query.
- https://docs.snowflake.com/en/sql-reference/sql/create-network-rule (fetched 2026-11-10)
- https://docs.snowflake.com/en/sql-reference/sql/create-secret (fetched 2026-11-10)
- https://docs.snowflake.com/en/sql-reference/sql/create-external-access-integration (fetched 2026-11-10)
- https://docs.snowflake.com/en/user-guide/security-access-control-privileges (fetched 2026-11-10)
- https://docs.snowflake.com/en/developer-guide/snowpark-container-services/service-network-communications (fetched 2026-11-10)
- https://docs.snowflake.com/en/developer-guide/snowpark-container-services/specification-reference (fetched 2026-11-10)
- https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-coding-agent (fetched 2026-11-10)
- https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-code-execution-tool (fetched 2026-11-10)
- https://docs.snowflake.com/en/user-guide/cortex-code/cloud-sandbox (fetched 2026-11-10)
- https://docs.snowflake.com/en/user-guide/cortex-code/cortex-code-snowsight/cloud-agents (fetched 2026-11-10)
- https://docs.snowflake.com/en/user-guide/cortex-code/cortex-code-automations (fetched 2026-11-10)
- https://docs.snowflake.com/en/sql-reference/sql/create-mcp-server (fetched 2026-11-10)
- https://docs.snowflake.com/en/sql-reference/parameters (COCO_CLOUD_AGENTS_NON_SNOWFLAKE_EGRESS_DISABLED; fetched 2026-11-10)
- Release notes (fetched 2026-11-10): Aug 26, 2026 Coding Agent GA; Aug 21, 2026 CoCo automations Preview; Oct 23, 2026 Coding agent MCP tool; Oct 28, 2026 Coding Agent thinking effort / stage mounts; Nov 2, 2026 CoCo automations GA