Clean, focused implementation that closes a genuine UX gap — the chat agent's "Active for" panel invisibility — without touching any activation, ledger, or SSE logic. The flag-follows-data pattern (ClassVar → AgentRegistration → Pydantic → TS type → frontend filter) is the right shape for this kind of capability surface, and the test suite is unusually thorough.

Issues

🟢 Minor — Multi-line comments on the new ClassVar exceed the one-line-max CLAUDE.md standard (agent.py:52–57, chat/agent.py:79–82)

The adjacent REQUIRED_CONNECTORS ClassVar in agent.py:49 is described in one line (# Empty list = no external connections required (the default for built-ins).). The new flag gets five. CLAUDE.md says "one short line max" for WHY comments.

src/gaia/agents/base/agent.py:52–57 — suggestion:

    # Registry reads this to include dynamic MCP consumers in the Settings "Active for" panel.
    CONSUMES_MCP_SERVERS: ClassVar[bool] = False

src/gaia/agents/chat/agent.py:79–82 — suggestion:

    # Dynamic MCP loader — registry exposes this for the Settings "Active for" panel.
    CONSUMES_MCP_SERVERS: ClassVar[bool] = True

Strengths

• Guard test is the right call. test_builtin_chat_registration_matches_class explicitly checks that the hardcoded registry value matches ChatAgent.CONSUMES_MCP_SERVERS — exactly the kind of sync risk a lazy factory creates, and exactly the right place to catch drift.
• MCP_DEFAULT_GRANT_SCOPES constant. Naming the canonical scope rather than inlining ['use'] is the correct move; the comment on it explains the "why" for dynamic consumers in one sentence.
• activatableAgents separation from relevantAgents. The frontend neatly keeps credential-grant eligibility (static required_connections) separate from activation eligibility (wider), and the connectorType === 'mcp_server' guard is baked into activatableAgents's ternary rather than duplicated at the render site.
• Test breadth. Base default, chat value, class↔registration sync, no-other-builtins regression guard, custom-agent round-trips (True and False), API serialization — these cover every layer the flag touches.

Verdict

Approve with suggestions. The two multi-line comment nits are the only findings; no bugs, no architectural concerns. Safe to merge after applying the one-liner suggestions (or skipping them — they're style-only).

Clean, well-scoped implementation. The design correctly separates the two inclusion paths (static REQUIRED_CONNECTORS vs. dynamic CONSUMES_MCP_SERVERS) without touching any activation, ledger, or SSE logic — exactly the right blast radius for a UI surfacing gap. One nit worth a one-line fix; everything else is solid.

Issues

🟢 Minor — Unnecessary defensive getattr in _reg_to_info (src/gaia/ui/routers/agents.py:115)

Every other field serialised by _reg_to_info uses direct attribute access (reg.min_memory_gb, reg.required_connections, reg.namespaced_agent_id). Since consumes_mcp_servers is now a proper dataclass field with a default, the getattr fallback is redundant and slightly misleading.

        consumes_mcp_servers=reg.consumes_mcp_servers,

🟢 Minor — Verbose comment block on AgentRegistration.consumes_mcp_servers (src/gaia/agents/registry.py:278–282)

GAIA convention (CLAUDE.md "Code Comments") is one short WHY line per field. The 5-line block mirrors what the field name and type already say; the neighbouring required_connections comment runs two lines. Trimming keeps the dataclass scannable.

    # Surfaced from the agent's CONSUMES_MCP_SERVERS ClassVar; True for dynamic MCP loaders.
    consumes_mcp_servers: bool = False

Same trimming applies to the parallel comment in src/gaia/ui/models.py:255–258 (4 lines; one line is enough for the same reason).

Strengths

• test_builtin_chat_registration_matches_class is the right kind of guard test: it catches exactly the failure mode (hardcoded flag drifts from the class definition) that the lazy-factory workaround creates, without coupling the test to internal module paths.
• Two-list design in ConnectorsSection.tsx (relevantAgents for per-credential grants, activatableAgents for the "Active for" panel) keeps the two sections independent and makes the intent immediately readable in the diff.
• MCP_DEFAULT_GRANT_SCOPES named constant eliminates a magic ['use'] string from the activation callback; the JSDoc on AgentActivationCard is updated in the same commit to explain the fallback path.
• TypeScript type is consumes_mcp_servers?: boolean (optional), so the frontend degrades gracefully against an older backend that doesn't send the field — the filter treats undefined as falsy and the agent simply won't appear.

Verdict

Approve. Both nits are stylistic and non-blocking. The implementation is correct, the guard test is smart, and the test matrix covers all four cases (base default, chat builtin, custom consumer, custom non-consumer) plus the serialisation round-trip.

@kovtcharov-amd -- My original commit
50fb140
for this PR got reported "All checks have passed" and Verdict "Approve with suggestions. The two multi-line comment nits are the only findings...".
So, I made suggested fixes in comments only, and this new commit somehow got Failed check:
Test GAIA CLI on Windows (Full Integration).
But it also got Verdict "Approve."
This failed test looks like a glitch.

Clean, narrow feature that closes a real UX gap in the connectors panel. One nit worth a one-line fix; everything else is solid. Approve with suggestion.

Summary

The "Active for" panel previously only listed agents that declare REQUIRED_CONNECTORS, so builtin:chat — which loads MCP servers dynamically and already honours the activation ledger — was invisible in the UI despite the backend working correctly. This PR closes the gap with a CONSUMES_MCP_SERVERS capability flag that flows exactly the same path as REQUIRED_CONNECTORS: agent class → AgentRegistration → AgentInfo (Python) → /api/agents → TypeScript AgentInfo. The implementation is narrow, the logic is correct, and the guard test that keeps the hardcoded builtin registration in sync with the class attribute is a smart pattern.

Issues Found

🟢 Minor — doc snippet missing from typing import ClassVar (docs/sdk/infrastructure/connectors.mdx:196–199)

The minimal snippet uses ClassVar[bool] but doesn't import it. The preceding GmailAgent example at line 147 does include the import, but a reader copying just this snippet will get a NameError.

```python
from typing import ClassVar
from gaia.agents.base.agent import Agent


class MyAgent(Agent):
    CONSUMES_MCP_SERVERS: ClassVar[bool] = True

---

## Strengths

- **Layered propagation is exact.** The flag follows the same `ClassVar → AgentRegistration → AgentInfo → TypeScript` pipeline as `REQUIRED_CONNECTORS`, making it easy for future contributors to reason about.
- **Guard test prevents silent drift.** `test_builtin_chat_registration_matches_class` directly asserts that the hardcoded `consumes_mcp_servers=True` in the lazy-factory builtin registration matches `ChatAgent.CONSUMES_MCP_SERVERS`. This is the right way to handle the necessary decoupling for import-cycle reasons.
- **Frontend changes are surgically scoped.** `activatableAgents` is a separate computed set that widens *only* the "Active for" block; `relevantAgents` and the per-agent credential grants section are untouched. The `MCP_DEFAULT_GRANT_SCOPES` constant eliminates the magic string in the toggle path. The `connectorType === 'mcp_server'` guard is correctly absorbed into `activatableAgents` (returns `[]` for non-MCP connectors) rather than duplicated at the render site.
- **Test coverage is thorough:** base default, chat flag, class↔registration consistency, "no other builtin sets the flag," custom-agent round-trip in both directions, `AgentInfo` Pydantic round-trip, and `_reg_to_info` serialization for a consumer vs. a non-consumer.

---

## Verdict

**Approve with suggestion** — one nit (missing import in the doc snippet). Safe to merge after applying; the doc fix is one line and can be done inline.

@kovtcharov-amd this PR is approved by @itomek. Can it be merged now?

@kovtcharov-amd this PR is approved by @itomek. Can it be merged now?

Yes, merged. Apologies for the delay.