Ollama Gateway: Supercharge your native Ollama with enterprise-grade API authentication, request auditing, rate limiting, and virtual model management—your secure, private AI gateway.
Default user name is admin@default.com and default password is Admin@123456!.
Native Ollama is great for personal use, but it lacks the enterprise features required for team collaboration and production deployment. OllamaGateway fills those gaps without changing your workflow.
| Feature | Native Ollama | OllamaGateway |
|---|---|---|
| Model Hosting & Inference | ✅ | ✅ |
| Multimodal | ✅ | ✅ |
| MCP | ✅ | ✅ |
| Function Call | ✅ | ✅ |
| Streaming | ✅ | ✅ |
| OpenAI API Translation | ✅ | |
| Anthropic API Translation | ✅ | ✅ |
| API Authentication (Bearer Token) | ❌ | ✅ |
| Multiple API Keys Management | ❌ | ✅ |
| Request & Response Auditing | ❌ | ✅ |
| API Rate Limiting | ❌ | ✅ |
| Virtual Model Overrides | ❌ | ✅ |
| Multi-backend Support | ❌ | ✅ |
| Load Balancing | ❌ | ✅ |
| Tiered Fallback | ❌ | ✅ |
| Default Model Support | ❌ | ✅ |
| Model Keep-alive (Ping) | ❌ | ✅ |
| Admin Management GUI | ❌ | ✅ |
| Chat/Embedding Segregation | ❌ | ✅ |
Try a running OllamaGateway here.
Run the installer with the desired port (omit it to choose an available port):
curl -fsSL https://raw.githubusercontent.com/aiursoftweb/ollamagateway/master/install.sh | sudo bash -s -- 8080Installation starts after three seconds; press Ctrl+C to cancel. Then open http://localhost:8080.
Requirements about how to run
- Install .NET 10 SDK and Node.js.
- Execute
npm installatwwwrootfolder to install the dependencies. - Execute
dotnet runto run the app. - Use your browser to view http://localhost:5000.
- Open the
.slnfile in the project path. - Press
F5to run the app.
First, install Docker here.
Then run the following commands in a Linux shell:
image=aiursoft/ollamagateway
appName=ollamagateway
sudo docker pull $image
sudo docker run -d --name $appName --restart unless-stopped -p 5000:5000 -v /var/www/$appName:/data $imageThat will start a web server at http://localhost:5000 and you can test the app.
The docker image has the following context:
| Properties | Value |
|---|---|
| Image | aiursoft/ollamagateway |
| Ports | 5000 |
| Binary path | /app |
| Data path | /data |
| Config path | /data/appsettings.json |
OllamaGateway has been officially tested and is fully compatible with popular ecosystem tools including Claude Code, Open-WebUI, Opencode, and Roocode. Its robust API translation makes it the ideal choice for deploying autonomous AI agents.
For the best experience with AI agents, we recommend the following models:
qwen3.6:27b-q8_0qwen3.6:35b-a3b-q4_K_M
OllamaGateway logs every request to a ClickHouse database. A pre-built Grafana dashboard template is provided at docs/grafana-config.json to visualize traffic, latency, and usage in real time.
- In Grafana, go to Dashboards → Import.
- Upload
docs/grafana-config.jsonor paste its contents. - When prompted, select your ClickHouse data source for the grafana-clickhouse-datasource input.
- Click Import.
Ollama Gateway exposes four client-facing dialects:
- Ollama native (
/api/chat) - OpenAI Chat Completions (
/v1/chat/completions) - OpenAI Responses (
/v1/responses) - Anthropic Messages (
/v1/messages)
Requests are decoded into a protocol-neutral representation and then encoded for the selected physical backend. Responses take the reverse path. Each physical-model backend records its actual protocol independently from its provider, so one OpenAI-compatible provider can host both Chat Completions and Responses backends.
When the client and physical backend use the same dialect, the gateway takes a transparent fast path: it preserves unknown request fields, response items, and stream events while changing only gateway-owned fields such as the model name. Cross-dialect calls use the neutral representation.
Responses supports text, image, custom-function, structured-output, reasoning, and streaming translation. Native Responses tools and stateful controls (previous_response_id, Conversations, stored responses, and background mode) are forwarded only to Responses-capable backends. If no native Responses route is available, those features return an explicit unsupported_feature error instead of being silently ignored.
The table below documents the pre-Responses Chat Completions, Ollama, and Anthropic parameter mappings:
- Paths ① and ③ enter via
/v1/chat/completionsand are handled byOpenAIController. - Paths ② and ④ enter via
/api/chatand are handled byProxyController. - Paths ⑤ and ⑥ enter via
/v1/messagesand are handled byAnthropicController. - OpenAI Responses enters via
/v1/responsesand is handled byOpenAIResponsesController; its mappings are described above rather than added to this legacy matrix.
"DB override" means a hard assignment (
=), not a null-coalescing assignment (??=). When the virtual model has a value configured in the database, the client-supplied value is always discarded, regardless of what the client sent.
| Parameter | ① OpenAI provider + OpenAI API |
② OpenAI provider + Ollama API |
③ Ollama provider + OpenAI API |
④ Ollama provider + Ollama API |
⑤ OpenAI provider + Anthropic API |
⑥ Ollama provider + Anthropic API |
|---|---|---|---|---|---|---|
| Temperature | ✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as "temperature" (passthrough) |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as "temperature" |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.temperature to Ollama |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.temperature to Ollama |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as "temperature" |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.temperature |
| Top P | ✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as "top_p" (passthrough) |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as "top_p" |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.top_p to Ollama |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.top_p to Ollama |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as "top_p" |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.top_p |
| Top K | ❌ Always discarded 🚫 OpenAI does not support it ➖ DB does not inject it |
❌ Always discarded ✏️ DB injects into options but the field is dropped before the OpenAI request is sent |
🚫 Client cannot pass it (no such field in OpenAI API) ✏️ DB injects if set ➡️ Forwarded as options.top_k to Ollama |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.top_k to Ollama |
❌ Always discarded 🚫 OpenAI does not support it |
🚫 Client cannot pass it ✏️ DB injects if set ➡️ Forwarded as options.top_k |
| Num Ctx | ❌ Always discarded 🚫 OpenAI does not support it ➖ DB does not inject it |
❌ Always discarded ✏️ DB injects into options but the field is dropped before the OpenAI request is sent |
🚫 Client cannot pass it (no such field in OpenAI API) ✏️ DB injects if set ➡️ Forwarded as options.num_ctx to Ollama |
✅ Client value arrives ✏️ DB hard-overrides if set ➡️ Forwarded as options.num_ctx to Ollama |
❌ Always discarded 🚫 OpenAI does not support it |
🚫 Client cannot pass it ✏️ DB injects if set ➡️ Forwarded as options.num_ctx |
| Thinking | Client setting is preserved; selected backend may override switch and effort | Client setting is preserved or translated; selected backend may override switch and effort | Client effort is translated; selected backend may force think on/off |
Client think is preserved; selected backend may force on/off |
Translated client setting; selected backend may override switch and effort | Translated client setting; selected backend may force think on/off |
| Images / Multimodal | ✅ Multimodal content array passes through as-is
➡️ Direct passthrough to OpenAI backend | ✅ Ollama images converted
➡️ images[] → content[text+image_url] | ✅ OpenAI content array extracted
➡️ content[image_url] → images[] | ✅ Images array passes through
➡️ Direct passthrough to Ollama | ❌ Not yet translated
🚫 Anthropic image blocks dropped | ❌ Not yet translated
🚫 Anthropic image blocks dropped |
Thinking overrides are configured per model backend. Ollama-native backends offer Do not override (default), Force on, and Force off. No override preserves the client's boolean or named think level; when absent, the upstream default applies. Cross-protocol requests translate the client control to the target protocol. OpenAI Chat Completions and Responses backends also offer an independent Do not override / Force on / Force off switch, sent as chat_template_kwargs.enable_thinking. This is a vLLM-style extension: leave the switch unconfigured for providers that do not support it, such as ordinary OpenAI backends. The adjacent question-mark buttons explain forwarding, field mappings, provider/model differences, and conflicting settings. These backends also offer a free-text reasoning effort override (for example, medium, high, or xhigh). Leave it empty to preserve the client setting or the upstream default when absent. The model must support the value you enter; Gateway does not map effort names to other levels. A configured depth also overrides an existing chat_template_kwargs.reasoning_effort so a client template argument cannot shadow it. The switch and depth are independent; for vLLM/Qwen3.8, for example, force the switch on and use xhigh when both should be fixed. Existing backends inherit the new switch without changing their depth settings.
Overrides apply to same-dialect direct forwarding and cross-dialect translation. After selecting a backend, Gateway applies that backend's configured value to the outgoing request: reasoning_effort for Chat Completions, reasoning.effort for Responses, or boolean think for Ollama. Every fallback attempt uses its own backend configuration; a failed attempt cannot leak its override into the next attempt.
On upgrade, existing virtual-model thinking settings move to each Ollama-native backend. OpenAI backends stop receiving the old virtual-model chat_template_kwargs.enable_thinking injection; clients that need it can supply it themselves or administrators can configure the new per-backend switch. A downgrade can restore a model-wide setting only when all of its backends have the same explicit override.
The application defaults are defined in src/Aiursoft.OllamaGateway/appsettings.json. The Docker image initializes /data/appsettings.json on its first start and keeps it for later starts; persist /data. Override nested values with double-underscore environment variables, such as Logging__LogLevel__Default.
ConnectionStringsselects the database provider and connection. Prepare the external database before choosing MySQL; keep the SQLite file persistent when using SQLite.Storageconfigures local application files. Keep the storage directory persistent and use the documented hostname and upload policies for shared downloads.AppSettingscontrols authentication and account synchronization. Choose local accounts or configure the OIDC authority, client and/signin-oidccallback; use environment variables for credentials.- Uncommented
GlobalSettingsentries override the database-backed administration settings. Leave entries commented to use the stored values or application defaults. Logging:LogLevelcontrols application and framework verbosity. Configure the optional ClickHouse exporter only when its server and credentials are ready.AllowedHostsis the ASP.NET Core Host-header allowlist, not a CORS setting.
There are many ways to contribute to the project: logging bugs, submitting pull requests, reporting issues, and creating suggestions.
Even if you with push rights on the repository, you should create a personal fork and create feature branches there when you need them. This keeps the main repository clean and your workflow cruft out of sight.
We're also interested in your feedback on the future of this project. You can submit a suggestion or feature request through the issue tracker. To make this process more effective, we're asking that these include more information to help define them more clearly.
Licensed under the MIT License.
Footnotes
-
Native Ollama partially supports OpenAI API. It lacks advanced system prompt mapping, parameter overriding, and role translation. ↩

{ "ConnectionStrings": { "AllowCache": "True", "DbType": "Sqlite", "DefaultConnection": "DataSource=app.db;Cache=Shared" // sudo docker run -d --name db -e MYSQL_RANDOM_ROOT_PASSWORD=true -e MYSQL_DATABASE=ollamagateway -e MYSQL_USER=ollamagateway -e MYSQL_PASSWORD=ollamagateway_password -p 3306:3306 hub.aiursoft.com/mysql //"DbType": "MySql", //"DefaultConnection": "Server=localhost;Database=ollamagateway;Uid=ollamagateway;Pwd=ollamagateway_password;" }, "Storage": { "Path": "/tmp/data", "PublicOrigin": "", "RequireDedicatedInlineOrigin": false, "AllowArbitraryInlineOnDedicatedOrigin": false, "SafeInlineMediaExtensions": ["svg"], "SafeInlineSvgSubfolders": ["project-logo"], "UploadPolicy": { "RequireAuthenticatedUser": false, "ReplaceSpacesWithHyphens": false } }, "AppSettings": { "AuthProvider": "Local", // OIDC or Local "DefaultRole": "", "PersistsSignIn": false, "OIDC": { // https://auth.aiursoft.com/application/o/ollamagateway "Authority": "https://auth.aiursoft.com/application/o/ollamagateway", "ClientId": "", "ClientSecret": "", "RolePropertyName": "groups", "UsernamePropertyName": "preferred_username", "UserDisplayNamePropertyName": "name", "EmailPropertyName": "email", "UserIdentityPropertyName": "sub" }, "Local": { "AllowRegister": true, "AllowWeakPassword": true } }, "GlobalSettings": { // You can initialize or override global settings here. // "Allow_User_Adjust_Nickname": "True", // "BrandName": "Aiursoft", // "BrandHomeUrl": "https://www.aiursoft.com/", // "Icp": "", // "ProjectName": "Ollama Gateway" }, "Logging": { "LogLevel": { "Default": "Information", "Microsoft.AspNetCore": "Warning" }, "Clickhouse": { "Enabled": false, "ConnectionString": "Host=localhost;Protocol=http;Port=8123;User=default;Password=;Database=AspNetCoreLogs", "TableName": "OllamaGatewayLogs" } }, "Clickhouse": { "Enabled": false, "ConnectionString": "Host=localhost;Protocol=http;Port=8123;User=default;Password=;Database=OllamaGatewayAudit", "TableName": "RequestLogs" }, "AllowedHosts": "*" }