diff --git a/config/installation_config.yml b/config/installation_config.yml
index 1d02405d2..e4892b526 100644
--- a/config/installation_config.yml
+++ b/config/installation_config.yml
@@ -184,7 +184,7 @@
 # MARK: Captain Config
 - name: CAPTAIN_LLM_PROVIDER
   display_title: 'Captain LLM Provider'
-  description: 'Qual provider o Captain usa: openai_api (padrão, API key tradicional) ou openai_codex_oauth (assinatura ChatGPT Plus via proxy interno).'
+  description: 'Qual provider o Captain usa: openai_api (padrão, API key tradicional), openai_codex_oauth (assinatura ChatGPT Plus via proxy interno) ou openai_hermes_gateway (Hermes Agent rodando como gateway HTTP local — ele faz o roteamento multi-modelo via OAuth).'
   value: 'openai_api'
   locked: false
 - name: CAPTAIN_CODEX_PROXY_URL
@@ -192,6 +192,21 @@
   description: 'URL base do proxy Codex interno quando CAPTAIN_LLM_PROVIDER=openai_codex_oauth. Default: http://localhost:3000/codex'
   value: 'http://localhost:3000/codex'
   locked: false
+- name: CAPTAIN_HERMES_GATEWAY_URL
+  display_title: 'Captain Hermes Gateway URL'
+  description: 'URL base do Hermes Gateway quando CAPTAIN_LLM_PROVIDER=openai_hermes_gateway. Default: http://host.docker.internal:9877 (Hermes rodando no host, container alcança via host.docker.internal).'
+  value: 'http://host.docker.internal:9877'
+  locked: false
+- name: CAPTAIN_HERMES_GATEWAY_MODEL
+  display_title: 'Captain Hermes Gateway Model'
+  description: 'Modelo a passar pro Hermes Gateway no formato <provider>/<model>. Default: anthropic/claude-opus-4-5. O Hermes faz o roteamento real e pode usar Codex/Anthropic/Gemini conforme config local em ~/.hermes/config.yaml.'
+  value: 'anthropic/claude-opus-4-5'
+  locked: false
+- name: CAPTAIN_HERMES_GATEWAY_API_KEY
+  display_title: 'Captain Hermes Gateway API Key (optional)'
+  description: 'API key opcional pro Hermes Gateway. Geralmente vazio (gateway local não exige auth). Se setado, vai no Authorization header das requisições do Captain pro Hermes.'
+  locked: false
+  type: secret
 - name: CAPTAIN_OPEN_AI_API_KEY
   display_title: 'OpenAI API Key'
   description: 'The API key used to authenticate requests to OpenAI services for Captain AI.'
diff --git a/enterprise/app/services/captain/llm/provider_config.rb b/enterprise/app/services/captain/llm/provider_config.rb
index ba84be7c3..f3004187a 100644
--- a/enterprise/app/services/captain/llm/provider_config.rb
+++ b/enterprise/app/services/captain/llm/provider_config.rb
@@ -9,9 +9,16 @@
 #   (CAPTAIN_CODEX_PROXY_URL, default http://localhost:3000/codex) e usa uma
 #   api_key dummy — o proxy ignora o Authorization header e usa OAuth interno.
 #
+# - openai_hermes_gateway: aponta para o Hermes Agent rodando em modo gateway
+#   (CAPTAIN_HERMES_GATEWAY_URL, default http://host.docker.internal:9877).
+#   O Hermes Gateway expõe API HTTP compatível com OpenAI e roteia internamente
+#   pra Codex/Anthropic/Gemini conforme sua config local em ~/.hermes/config.yaml.
+#   Auth: usa CAPTAIN_HERMES_GATEWAY_API_KEY se setado, senão dummy (gateway local).
+#
 # O "legacy" ruby-openai usado para PDF/Files API NÃO deve usar esse módulo:
-# o endpoint Codex não expõe Files API, então esses serviços continuam
-# apontando sempre para OpenAI tradicional.
+# o endpoint Codex/Hermes não expõe Files API nem /embeddings, então esses
+# serviços continuam apontando sempre para OpenAI tradicional via
+# legacy_openai_settings.
 class Captain::Llm::ProviderConfig
   DEFAULT_MODEL = 'gpt-4.1-mini'.freeze
   DEFAULT_OPENAI_ENDPOINT = 'https://api.openai.com'.freeze
@@ -23,12 +30,20 @@ class Captain::Llm::ProviderConfig
   # endpoint Codex da OpenAI via ChatGPT Plus.
   DEFAULT_CODEX_MODEL = 'gpt-5.2'.freeze
 
+  # Hermes Gateway: defaults para o setup standard do Hermes Agent rodando
+  # como gateway HTTP local. O gateway escuta em 0.0.0.0:9877 por padrão e
+  # aceita o nome do modelo no formato `<provider>/<model>`.
+  DEFAULT_HERMES_GATEWAY_URL = 'http://host.docker.internal:9877'.freeze
+  DEFAULT_HERMES_GATEWAY_MODEL = 'anthropic/claude-opus-4-5'.freeze
+  HERMES_GATEWAY_DUMMY_KEY = 'hermes-gateway'.freeze
+
   # Modelo leve pra tasks de background (extração de memória, verificação de
-  # contradição, traduções internas). Quando usamos Codex, reutilizamos o
-  # mesmo modelo do chat — o endpoint não expõe gpt-4o-mini.
+  # contradição, traduções internas). Quando usamos Codex/Hermes, reutilizamos
+  # o mesmo modelo do chat — esses endpoints não expõem gpt-4o-mini.
   LIGHT_MODEL_DEFAULTS = {
     'openai_api' => 'gpt-4o-mini',
-    'openai_codex_oauth' => DEFAULT_CODEX_MODEL
+    'openai_codex_oauth' => DEFAULT_CODEX_MODEL,
+    'openai_hermes_gateway' => DEFAULT_HERMES_GATEWAY_MODEL
   }.freeze
 
   class << self
@@ -40,12 +55,16 @@ class Captain::Llm::ProviderConfig
       provider == 'openai_codex_oauth'
     end
 
+    def hermes_gateway?
+      provider == 'openai_hermes_gateway'
+    end
+
     # Retorna { api_key:, api_base:, model: } para RubyLLM/Agents.
     def settings
-      if codex_oauth?
-        codex_settings
-      else
-        openai_api_settings
+      case provider
+      when 'openai_codex_oauth'    then codex_settings
+      when 'openai_hermes_gateway' then hermes_gateway_settings
+      else                              openai_api_settings
       end
     end
 
@@ -88,6 +107,14 @@ class Captain::Llm::ProviderConfig
       }
     end
 
+    def hermes_gateway_settings
+      {
+        api_key: cfg('CAPTAIN_HERMES_GATEWAY_API_KEY').presence || HERMES_GATEWAY_DUMMY_KEY,
+        api_base: (cfg('CAPTAIN_HERMES_GATEWAY_URL').presence || DEFAULT_HERMES_GATEWAY_URL).chomp('/'),
+        model: cfg('CAPTAIN_HERMES_GATEWAY_MODEL').presence || DEFAULT_HERMES_GATEWAY_MODEL
+      }
+    end
+
     def openai_api_settings
       {
         api_key: cfg('CAPTAIN_OPEN_AI_API_KEY'),
diff --git a/spec/enterprise/services/captain/llm/provider_config_spec.rb b/spec/enterprise/services/captain/llm/provider_config_spec.rb
index 5e9ee518e..45dbd463e 100644
--- a/spec/enterprise/services/captain/llm/provider_config_spec.rb
+++ b/spec/enterprise/services/captain/llm/provider_config_spec.rb
@@ -63,5 +63,72 @@ RSpec.describe Captain::Llm::ProviderConfig do
         expect(described_class.settings[:api_base]).to eq(described_class::DEFAULT_CODEX_PROXY_URL)
       end
     end
+
+    context 'when provider is openai_hermes_gateway' do
+      before do
+        InstallationConfig.create!(name: 'CAPTAIN_LLM_PROVIDER', value: 'openai_hermes_gateway')
+        InstallationConfig.create!(name: 'CAPTAIN_HERMES_GATEWAY_URL', value: 'http://host.docker.internal:9877')
+        InstallationConfig.create!(name: 'CAPTAIN_HERMES_GATEWAY_MODEL', value: 'anthropic/claude-opus-4-5')
+      end
+
+      it 'returns the gateway URL with dummy api_key when no key is configured' do
+        settings = described_class.settings
+        expect(settings[:api_key]).to eq(described_class::HERMES_GATEWAY_DUMMY_KEY)
+        expect(settings[:api_base]).to eq('http://host.docker.internal:9877')
+        expect(settings[:model]).to eq('anthropic/claude-opus-4-5')
+      end
+
+      it 'honors CAPTAIN_HERMES_GATEWAY_API_KEY when present' do
+        InstallationConfig.create!(name: 'CAPTAIN_HERMES_GATEWAY_API_KEY', value: 'sk-hermes-real')
+        expect(described_class.settings[:api_key]).to eq('sk-hermes-real')
+      end
+
+      it 'honors a custom CAPTAIN_HERMES_GATEWAY_MODEL value' do
+        InstallationConfig.find_by!(name: 'CAPTAIN_HERMES_GATEWAY_MODEL').update!(value: 'openai/gpt-5.4')
+        expect(described_class.settings[:model]).to eq('openai/gpt-5.4')
+      end
+
+      it 'reports hermes_gateway? as true and codex_oauth? as false' do
+        expect(described_class.hermes_gateway?).to be true
+        expect(described_class.codex_oauth?).to be false
+      end
+
+      it 'strips trailing slash from gateway URL' do
+        InstallationConfig.find_by!(name: 'CAPTAIN_HERMES_GATEWAY_URL').update!(value: 'http://host.docker.internal:9877/')
+        expect(described_class.settings[:api_base]).to eq('http://host.docker.internal:9877')
+      end
+
+      it 'uses default model when CAPTAIN_HERMES_GATEWAY_MODEL is missing' do
+        InstallationConfig.find_by!(name: 'CAPTAIN_HERMES_GATEWAY_MODEL').delete
+        expect(described_class.settings[:model]).to eq(described_class::DEFAULT_HERMES_GATEWAY_MODEL)
+      end
+
+      it 'uses default URL when CAPTAIN_HERMES_GATEWAY_URL is missing' do
+        InstallationConfig.find_by!(name: 'CAPTAIN_HERMES_GATEWAY_URL').delete
+        expect(described_class.settings[:api_base]).to eq(described_class::DEFAULT_HERMES_GATEWAY_URL)
+      end
+    end
+
+    context 'when default provider (openai_api) is in use' do
+      it 'reports hermes_gateway? as false' do
+        expect(described_class.hermes_gateway?).to be false
+      end
+    end
+
+    describe '.light_model' do
+      it 'returns gpt-4o-mini for openai_api' do
+        expect(described_class.light_model).to eq('gpt-4o-mini')
+      end
+
+      it 'returns DEFAULT_CODEX_MODEL for openai_codex_oauth' do
+        InstallationConfig.create!(name: 'CAPTAIN_LLM_PROVIDER', value: 'openai_codex_oauth')
+        expect(described_class.light_model).to eq(described_class::DEFAULT_CODEX_MODEL)
+      end
+
+      it 'returns DEFAULT_HERMES_GATEWAY_MODEL for openai_hermes_gateway' do
+        InstallationConfig.create!(name: 'CAPTAIN_LLM_PROVIDER', value: 'openai_hermes_gateway')
+        expect(described_class.light_model).to eq(described_class::DEFAULT_HERMES_GATEWAY_MODEL)
+      end
+    end
   end
 end