Cercloud
Managed shared skills
An administrator publishes text-only skills alongside a model policy. The host verifies and activates the complete bundle, then gives each new turn a fixed copy of its assigned skills. Publishing an update does not change instructions halfway through work. The next turn receives the new versions. Removing a skill removes it from the next turn's catalog; rolling back a publication restores the assigned older version in a new policy revision.
What developers and agents see
Shared skills have IDs such as enterprise/ORGANIZATION-ID/review. They cannot
replace built-in skills or workflow protocols, even if they have the same title.
The catalog identifies the source as enterprise and includes the assigned
version. Built-in entries identify their source as builtin and version as
bundled.
The main agent sees each assigned skill's ID, title, description and version in
its turn instructions. It reads the full text when needed with get_shared_skill.
Delegated agents inherit the same snapshot and receive the same read-only tool.
The tool returns text and metadata; it does not run scripts, install files, or
grant additional model or tool permissions. Personal skill files are unchanged.
The turn instructions explicitly replace earlier skill assignments in conversation history. Old conversation text is not deleted. When no shared skills are assigned, the prompt says so instead of silently leaving an earlier assignment in effect. As with other written instructions, a model's compliance with skill prose is not a deterministic security boundary. Model restrictions remain enforced in code.
External MCP clients can use cercano_skills with action: "list" or
action: "get" and the exact namespaced ID. Each request reads the host's current
verified bundle and includes source/version metadata. An MCP client controls its
own outer conversation; Cercano does not claim to pin that client's entire turn.
Host and worker behavior
A managed worker receives the complete pinned policy and skills over the private host-worker stream. It checks assignment IDs, versions, byte lengths and content digests before building providers or running tools. Enterprise credentials never enter this snapshot. A dedicated settings-aware RPC prevents an older worker from silently ignoring the new fields. Missing or incomplete settings stop the turn.
The snapshot does not authorize model requests. Every physical model attempt, including retries, still asks the host's current policy authority. Revocation or expiry can stop work even when its pinned instructions are older.
The existing signed-bundle download/cache logic remains responsible for trust, complete activation, restart behavior and synchronization acknowledgements. Skill discovery also requires a usable managed connection, so revoked members cannot continue reading cached assignments through the catalog RPCs.
Verification and remaining work
Automated tests exercise signed updates/removal/rollback, snapshot mutation isolation, cross-organization lookup, source/version discovery, MCP formatting, main-worker tool execution, delegated-agent tool execution, malformed bundles, and rejection of workers that only support the earlier authorization protocol. All use disposable local data; they do not contact Google or real model services.
Managed routing now applies administrator defaults and ordered fallbacks. The administration interface and reproducible two-company workflow are implemented; the workflow exercises shared-skill retrieval through an actual agent tool loop. Live Google and Keychain acceptance remain pending. See the implementation status.