Home | Chopingisht Ollama Gateway

Enterprise Ready

Security and Auditing by Design.

Bearer Token Auth

Create multiple API keys for different users and applications, each with its own fine-grained permissions.

Clickhouse Auditing

Detailed auditing for every request and response, stored in high-performance Clickhouse storage for compliance.

Team Management

A built-in Role-Based Access Control (RBAC) system for managing your team's access to AI models.

Model Keep-alive

Periodically pings underlying models to ensure they stay loaded in memory for instant response times.

Default Model Support

Automatically redirect requests to a default virtual model if no model is specified in the API call.

Native Passthrough

Seamless support for native Ollama features including MCP tools, embedding, images, and stream mode.

Agent & Ecosystem Ready

Perfect for Agent Deployment

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.

Recommended Models

qwen3.6:27b-q8_0
qwen3.6:35b-a3b-q4_K_M

Native Power, AI Development

Supercharge AI Development with Claude Code, Opencode & Qwen Code

OllamaGateway seamlessly connects your local Ollama capabilities to professional AI development tools. Whether you are using Claude Code, Opencode or Qwen Code, you can now leverage your own hardware for automated coding and development tasks—all verified and officially supported by our team.

Verified compatibility with Claude Code agent
Optimized for Opencode & Qwen Code ecosystems
Secure local inference with zero data leakage

Opencode and Qwen Code working with OllamaGateway

Cross-Protocol Compatibility

Inspect one translation route at a time.

Choose the API your client sends on the left and the protocol spoken by the physical backend on the right. The active route shows exactly what is preserved, translated, rejected, and overridden.

01Requester's API

Active route 01 / 12

Client Ollama Chat /api/chat

Backend Ollama Native /api/chat

Native / same dialect

The gateway keeps the protocol shape and rewrites only gateway-owned values and configured policy fields.

02Physical backend protocol

Provider type and backend protocol are separate. No physical Anthropic backend adapter exists today.

Current route contract

Ollama → Ollama

Only the selected route is shown

Implementation method

Preserved

Mapped / reconstructed

Discarded, rejected, or preferred

Database override

Direction	Implementation method	Preserved	Mapped / reconstructed	Discarded, rejected, or preferred	Database override
Ollama Chat client routes `/api/chat` → 3 backend protocols
`/api/chat` Ollama → Ollama Transparent	`OllamaChatRequestDecoder` performs capability inspection; `OllamaChatProviderRequestAdapter` clones the original JSON; `OllamaChatWireResponseWriter` keeps NDJSON records in the Ollama dialect.	Unknown request fields, `messages`, `tools`, `options`, `format`, and unknown JSON response fields are retained. Malformed upstream NDJSON records are forwarded raw.	Only gateway-owned fields are rewritten: virtual model → physical model upstream, physical model → virtual model downstream; `stream` is normalized; missing `keep_alive` receives the request/provider default.	No known Ollama field is intentionally dropped on this route. Unknown fields make an Ollama-native backend preferred; if none is available, the request may use a documented translated subset. Valid JSON may be reserialized.	`temperature`, `top_p`, `top_k`, `num_predict`, `num_ctx`, `repeat_penalty`, and `think` are hard-overridden when configured.
`/api/chat` Ollama → OpenAI Chat Translated	`OllamaChatRequestDecoder` → `GatewayChatRequest` → `OpenAiChatProviderRequestAdapter`. The response reverses through `OpenAiChatProviderResponseDecoder` and `OllamaChatClientResponseWriter`.	Message roles/text, function definitions, function calls/results, streaming intent, temperature, top-p, token limit, images, structured format, and token usage survive when representable.	`images[]` → OpenAI `image_url` data parts; `num_predict` → `max_tokens`; `think` → `chat_template_kwargs.enable_thinking`; Ollama tool calls and responses are rebuilt as OpenAI tools and back again.	`top_k`, `num_ctx`, `repeat_penalty`, and `keep_alive` have no OpenAI Chat mapping. Unknown Ollama controls prefer a native backend but are omitted when this translation path is the only eligible route.	OpenAI target fields `temperature`, `top_p`, and `max_tokens` use VM values when set. VM `Thinking` overrides the translated client value.
`/api/chat` Ollama → Responses Translated	`OllamaChatRequestDecoder` → canonical messages/items → `OpenAiResponsesProviderRequestAdapter`. Typed Responses Items/events are decoded and rebuilt as Ollama JSON/NDJSON.	Roles, text, system/developer instructions, images, function tools, function calls/results, temperature, top-p, token limit, structured format, stream, reasoning output, finish state, and usage.	System messages → `instructions`; messages → Responses `input` Items; images → `input_image`; `num_predict` → `max_output_tokens`; `format` → `text.format`; function tool shapes are flattened.	Ollama `top_k`, `num_ctx`, `repeat_penalty`, `keep_alive`, and boolean `think` are not encoded into Responses. Responses-native tools cannot be requested from the Ollama dialect.	VM `Temperature`, `TopP`, and `NumPredict` override `temperature`, `top_p`, and `max_output_tokens`. Other VM sampling fields are not applicable.
OpenAI Chat Completions client routes `/v1/chat/completions` → 3 backend protocols
`/v1/chat/completions` OpenAI Chat → Ollama Translated	`OpenAiChatRequestDecoder` → `GatewayChatRequest` → `OllamaChatProviderRequestAdapter`. Ollama response records are decoded into canonical events and emitted as Chat Completion JSON/SSE.	Roles/text, data-URI images, function tools/calls/results, temperature, top-p, token limit, reasoning text, structured output intent, stream, finish reason, and usage.	OpenAI content parts → Ollama `content` + `images[]`; `max_tokens`/`max_completion_tokens` → `num_predict`; `reasoning_effort` → boolean `think`; `response_format` → Ollama `format`.	Remote image URLs have no portable Ollama-native representation; base64/data-URI is the verified conversion. Chat-only controls such as penalties/logprobs/seed prefer a Chat backend and are omitted when Ollama is the only eligible route.	All Ollama VM controls apply: `temperature`, `top_p`, `top_k`, `num_predict`, `num_ctx`, `repeat_penalty`, and `think`. Ollama-only values may therefore be injected even though the client dialect cannot express them.
`/v1/chat/completions` OpenAI Chat → OpenAI Chat Transparent	`OpenAiChatRequestDecoder` inspects capabilities; the provider adapter clones the source JSON; `OpenAiChatWireResponseWriter` forwards Chat JSON/SSE while observing text, reasoning, tools, and usage.	Unknown request fields, message/content extensions, `response_format`, `reasoning_effort`, vendor fields, unknown response fields, null values, and unrecognized SSE lines remain in the Chat dialect.	Model identity is masked in both directions; `stream` is normalized; streaming requests merge `stream_options.include_usage=true`; null message content is normalized to an empty string.	No known Chat field is intentionally dropped on this route. Unknown/Chat-only fields prefer a Chat Completions backend, with semantic translation as fallback when no matching backend is available. Valid JSON may be reserialized.	VM `Temperature`, `TopP`, `NumPredict`, and `Thinking` override their target fields. Existing nested `chat_template_kwargs` members are preserved around `enable_thinking`.
`/v1/chat/completions` OpenAI Chat → Responses Translated	`OpenAiChatRequestDecoder` → canonical IL → `OpenAiResponsesProviderRequestAdapter`. Responses output Items/events return through the Responses decoder and Chat client writer.	Roles/text, images, function tools/calls/results, temperature, top-p, maximum output tokens, reasoning effort/output, structured output, streaming, finish reason, and usage.	System/developer messages → `instructions`; messages → `input` Items; function schemas are flattened; named `tool_choice.function.name` → Responses `tool_choice.name`; `max_completion_tokens` → `max_output_tokens`; `response_format` → `text.format`.	`chat_template_kwargs.enable_thinking` has no Responses mapping. Chat-only controls prefer a Chat backend and are omitted on fallback. Responses annotations, citations, and noncanonical output Items cannot be reconstructed in a Chat Completion response.	VM `Temperature`, `TopP`, and `NumPredict` override Responses target fields. VM `Thinking`, `TopK`, `NumCtx`, and `RepeatPenalty` are not Responses controls.
OpenAI Responses client routes `/v1/responses` → 3 backend protocols · native state passthrough
`/v1/responses` Responses → Ollama Translated subset	`OpenAiResponsesRequestDecoder` → canonical IL → `OllamaChatProviderRequestAdapter`. Ollama output becomes typed Responses Items and typed SSE events.	Instructions, message Items, text, images, function calls/results, function tool definitions, temperature, top-p, output-token limit, reasoning effort/output, structured format, stream, finish reason, and usage.	`instructions` → system message; `input_text`/`input_image` → Ollama content/images; `max_output_tokens` → `num_predict`; `reasoning.effort` → `think`; `text.format` → `format`; response events are synthesized with stable output indexes.	`previous_response_id`, `conversation`, `store:true`, `background:true`, native tools, and unknown Items require a Responses backend. Stateless representable requests fall back to translation.	All Ollama VM controls apply and override translated values: sampling, context, repeat penalty, token limit, and thinking. Ollama-only DB controls may be injected.
`/v1/responses` Responses → OpenAI Chat Translated subset	`OpenAiResponsesRequestDecoder` → canonical IL → `OpenAiChatProviderRequestAdapter`. Chat completions/chunks are decoded and emitted as Responses JSON/typed SSE.	Instructions, text/images, function tools/calls/results, temperature, top-p, token limit, reasoning effort/output, structured output, stream, finish state, and usage.	Input Items → Chat `messages`; instructions → system message; Responses function tool/choice shapes → Chat wrappers; `max_output_tokens` → `max_tokens`; `reasoning.effort` → `reasoning_effort`; `text.format` → `response_format`.	State/background requests require a Responses backend. Native Responses tools and unknown Items also require Responses. Metadata and unknown top-level controls prefer Responses and are omitted on fallback. Responses-only annotations/citations cannot be recreated from Chat output.	VM `Temperature`, `TopP`, `NumPredict`, and `Thinking` override Chat target controls; Chat-only VM thinking is injected through `chat_template_kwargs`.
`/v1/responses` Responses → Responses Transparent, state-aware	The request adapter deep-clones the original Responses JSON. `OpenAiResponsesWireResponseWriter` preserves unknown output Items and SSE event names/data/comments/id/retry without passing the payload through canonical IL.	Unknown top-level and nested request fields, native tools, metadata, future Items, annotations, citations, unknown response fields, future SSE event types, and JSON/non-JSON event data are retained.	Only model identity, stream intent, and configured VM options are gateway-owned. Upstream physical model names are replaced with the virtual model in response objects and events.	`previous_response_id`, `conversation`, `background`, and `store` retain their native Responses semantics. This is not byte-transparent because JSON may be reserialized.	VM `Temperature`, `TopP`, and `NumPredict` override `temperature`, `top_p`, and `max_output_tokens`. Other nested Responses fields remain untouched.
Anthropic Messages client routes `/v1/messages` → 3 backend protocols · no physical Anthropic adapter
`/v1/messages` Anthropic → Ollama Translated	`AnthropicChatRequestDecoder` → canonical IL → `OllamaChatProviderRequestAdapter`. Ollama JSON/NDJSON becomes Anthropic message JSON or `message_start/content_block_*/message_delta/message_stop` SSE.	System text, message text, base64 images, tool definitions, `tool_use`/`tool_result`, tool choice modes, temperature, top-p, max tokens, visible reasoning text, stream, finish reason, and usage.	Anthropic image sources → Ollama `images[]`; `max_tokens` → `num_predict`; `tool_use` → function call; `tool_result` → tool message; reasoning text → Ollama thinking; Ollama output is rebuilt as Anthropic content blocks.	Remote image URLs are not a portable Ollama-native contract; base64 is the verified path. Thinking signatures are discarded. Unknown content blocks remain a hard rejection; unknown top-level controls are omitted when this translated fallback is used.	All Ollama VM controls apply. `TopK`, `NumCtx`, `RepeatPenalty`, and `Thinking` can be injected from DB even though Anthropic request fields do not directly express them.
`/v1/messages` Anthropic → OpenAI Chat Translated	`AnthropicChatRequestDecoder` → canonical IL → `OpenAiChatProviderRequestAdapter`. Chat JSON/SSE is decoded and reconstructed as Anthropic Messages output.	System/messages, text, base64 and URL images, function tools/calls/results, tool choice, temperature, top-p, max tokens, reasoning text, stream, finish reason, and usage.	Anthropic system content → system message; image blocks → OpenAI `image_url`; `input_schema` → function `parameters`; `tool_use`/`tool_result` → Chat tool calls/messages; `max_tokens` remains the Chat token limit.	Thinking signatures are discarded. Unknown content blocks remain a hard rejection. Anthropic-only top-level controls prefer a future Anthropic backend but are omitted when Chat is the only eligible route.	VM `Temperature`, `TopP`, `NumPredict`, and `Thinking` override Chat target fields. Ollama-only DB sampling/context controls do not apply to Chat.
`/v1/messages` Anthropic → Responses Translated	`AnthropicChatRequestDecoder` → canonical IL → `OpenAiResponsesProviderRequestAdapter`. Responses Items/events are decoded and rebuilt as Anthropic content blocks and SSE.	System instructions, text, base64/URL images, function tools/calls/results, tool choice, temperature, top-p, max tokens, response reasoning text, stream, finish state, and usage.	System → `instructions`; messages/images → Responses input Items; Anthropic tools → Responses function tools; named tool choice → Responses `name`; `max_tokens` → `max_output_tokens`; Responses reasoning/text/function output → Anthropic blocks.	Anthropic input thinking blocks and signatures are not encoded into Responses input. Unknown content blocks cannot use this route; unknown top-level controls are omitted on fallback. Native Responses-only output Items, annotations, and citations cannot be represented in Anthropic output.	VM `Temperature`, `TopP`, and `NumPredict` override Responses fields. VM `Thinking`, `TopK`, `NumCtx`, and `RepeatPenalty` have no Responses target field.

Configuration precedence

Database values are hard assignments

A configured virtual-model value replaces the client-derived value after translation. An unset value leaves the translated value unchanged.

OllamaTemperature · TopP · TopK · NumPredict · NumCtx · RepeatPenalty · Thinking

OpenAI ChatTemperature · TopP · NumPredict · Thinking

ResponsesTemperature · TopP · NumPredict

OpenAI Responses

Native state passthrough

Stateful controls pass through to native Responses providers and are never silently translated.

previous_response_idconversationstore:truebackground:true

→ requires a Responses-capable provider

Embedding route inspector

Two client dialects × two Provider types

Embedding adapters select by ProviderType, not chat BackendProtocol

01Client endpoint

Embedding route 01 / 04

ClientOpenAI Embeddings/v1/embeddings

ProviderOpenAI-compatible/v1/embeddings

Native / same dialect

02Backend Provider type

OpenAI Embeddings → OpenAI-compatible

Current route only

Preserved / mapped

Discarded / restriction

Policy and implementation

Narrow operation

Ollama Generate

Ollama only

POST /api/generate

Ollama JSON/NDJSON stays in the native dialect; model identity, answer, finish data, and usage are observed and rewritten.

No Chat Completions or Responses translation. A non-Ollama backend returns 501.

Gateway-owned operation

Model discovery

No upstream call

/v1/models*/api/tags/api/ps/api/version

Virtual-model names and gateway-owned metadata are rendered directly in the requested client shape.

Physical provider metadata is not proxied; Ollama runtime details are placeholders or omitted.

OpenAI embedding input	OpenAI-type Provider	The original request and successful response JSON are cloned; unknown OpenAI embedding fields are preserved; only model identity is rewritten.	No chat protocol translation occurs. Backend `Protocol` is not consulted for embeddings.	`OpenAiEmbeddingClientAdapter` + `OpenAiEmbeddingProviderAdapter`. No VM sampling/context override is injected into an OpenAI embedding request.
OpenAI embedding input	Ollama-type Provider	OpenAI `input` is sent as Ollama `input`; `embeddings[]` and `prompt_eval_count` are reconstructed as OpenAI `data[]` and `usage`.	OpenAI embedding extensions outside the shared input are not copied into the cross-dialect request or reconstructed response.	VM `NumCtx`, `Temperature`, `TopP`, `TopK`, and `RepeatPenalty` are injected into Ollama `options`.
Ollama embedding input	OpenAI-type Provider	Ollama `input` or legacy `prompt` becomes OpenAI `input`; OpenAI `data[].embedding` is reconstructed as Ollama `embeddings[]`.	Ollama-only embedding fields such as `options`, `truncate`, `dimensions` variants, and `keep_alive` are not copied into the OpenAI request unless represented by the shared input.	Provider adapter selection uses `ProviderType`, not chat `BackendProtocol`. No Ollama VM options are injected into an OpenAI embedding target.
Ollama embedding input	Ollama-type Provider	The original request and successful response JSON are cloned; unknown fields are preserved; model identity is rewritten.	The provider-level `keep_alive` replaces the client value on this route.	VM `NumCtx`, `Temperature`, `TopP`, `TopK`, and `RepeatPenalty` hard-override Ollama `options`.
`POST /api/generate`	Ollama Native only	Ollama generate JSON/NDJSON is forwarded in the same dialect; model identity, answer text, finish data, and usage are observed and the virtual model name is restored.	There is no Chat Completions or Responses translation. A non-Ollama backend returns `501`.	VM Thinking, NumCtx, Temperature, TopP, TopK, NumPredict, and RepeatPenalty override the request; missing `keep_alive` uses the Provider value.
Model discovery endpoints	Gateway-generated response	Virtual-model names and gateway-owned metadata are rendered directly in the requested client shape.	These endpoints do not proxy physical provider metadata. Ollama size/digest/runtime details are placeholders or omitted.	This avoids recursive upstream discovery when a Provider points back to the gateway and keeps the public catalog scoped to virtual models.

High Availability

Load Balancing, Tiered Fallback & Mixed Providers

OllamaGateway goes beyond simple proxying. It intelligently distributes requests across multiple underlying models, automatically falls back when a provider is unavailable, and lets you combine models from different providers into a single virtual model.

Smart Load Balancing

Distribute API requests across multiple underlying Ollama instances or OpenAI-compatible providers. Configure weights and priorities for each backend to optimize throughput and resource utilization.

Tiered Fallback

Define fallback chains for your models. If the primary provider fails or times out, OllamaGateway automatically steps down to the next available backend — ensuring maximum uptime without manual intervention.

Multi-Provider Virtual Models

Blend multiple models from entirely different providers into a single virtual model. Mix Ollama backends with OpenAI-compatible services — your clients see one unified endpoint while the gateway handles all routing and translation behind the scenes.

Native vs. Gateway

Why choose OllamaGateway?

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 Comparison	Native Ollama	OllamaGateway
Model Hosting & Inference
Multimodal
MCP
Function call
Streaming
OpenAI API Translation
Anthropic API Translation
API Authentication (Bearer)
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

Ready to secure your AI gateway?

Go to Dashboard Deployment Guide View on GitHub

Supercharge Your Native Ollama.

Security and Auditing by Design.

Bearer Token Auth

Clickhouse Auditing

Team Management

Model Keep-alive

Default Model Support

Native Passthrough

Intelligent
Virtual Models

Chat Models

Embedding Models

Parameter Overrides

Perfect for Agent Deployment

Recommended Models

Supercharge AI Development with Claude Code, Opencode & Qwen Code

Inspect one translation route at a time.

Ollama → Ollama

Preserved

Mapped / reconstructed

Discarded, rejected, or preferred

Database override

Database values are hard assignments

Native state passthrough

Two client dialects × two Provider types

OpenAI Embeddings → OpenAI-compatible

Preserved / mapped

Discarded / restriction

Policy and implementation

Ollama Generate

Model discovery

Load Balancing, Tiered Fallback & Mixed Providers

Smart Load Balancing

Tiered Fallback

Multi-Provider Virtual Models

Why choose OllamaGateway?

Ready to secure your AI gateway?

Supercharge Your Native Ollama.

Security and Auditing by Design.

Bearer Token Auth

Clickhouse Auditing

Team Management

Model Keep-alive

Default Model Support

Native Passthrough

Intelligent Virtual Models

Chat Models

Embedding Models

Parameter Overrides

Perfect for Agent Deployment

Recommended Models

Supercharge AI Development with Claude Code, Opencode & Qwen Code

Inspect one translation route at a time.

Ollama → Ollama

Preserved

Mapped / reconstructed

Discarded, rejected, or preferred

Database override

Database values are hard assignments

Native state passthrough

Two client dialects × two Provider types

OpenAI Embeddings → OpenAI-compatible

Preserved / mapped

Discarded / restriction

Policy and implementation

Ollama Generate

Model discovery

Load Balancing, Tiered Fallback & Mixed Providers

Smart Load Balancing

Tiered Fallback

Multi-Provider Virtual Models

Why choose OllamaGateway?

Ready to secure your AI gateway?

Intelligent
Virtual Models