oocla

README

oocla

日本語

An HTTP server that speaks both the Ollama API and the OpenAI API, backed by the claude CLI. Clients of either API can request models by names like opus / sonnet / haiku. The name is OpenAI + Ollama + CLAude.

$ oocla serve
oocla listening on 127.0.0.1:11434

$ curl localhost:11434/api/chat -d '{
    "model": "haiku",
    "stream": false,
    "messages": [{"role": "user", "content": "Capital of Japan?"}]
  }'
{"model":"haiku:latest","message":{"role":"assistant","content":"Tokyo"},
 "done":true,"done_reason":"stop","prompt_eval_count":170,"eval_count":81}

$ curl localhost:11434/v1/chat/completions -d '{
    "model": "haiku",
    "messages": [{"role": "user", "content": "Capital of Japan?"}]
  }'
{"id":"chatcmpl-1","object":"chat.completion","model":"haiku:latest",
 "choices":[{"index":0,"message":{"role":"assistant","content":"Tokyo"},"finish_reason":"stop"}], ...}

Principles

• Tools are Ollama's JSON tool definitions only. Claude Code's built-in tools are all disabled
• Claude Code's default system prompt is disabled; the model behaves as a plain LLM
• No involvement in authentication. The environment's claude CLI is used with whatever credentials it already has
• No conversation is ever stored: every turn runs with --no-session-persistence
• Standard library only. No external dependencies

Requirements

An authenticated claude CLI on PATH. That is all. oocla itself is a single binary with no runtime dependencies; Go is not needed.

Install

Homebrew (macOS or Linux):

$ brew install kazufusa/tap/oocla

With a Go toolchain:

$ go install github.com/kazufusa/oocla/cmd/oocla@latest

Or download oocla_<version>_<os>_<arch> from the releases page and unpack it.

$ tar -xzf oocla_1.2.0_linux_amd64.tar.gz
$ ./oocla_1.2.0_linux_amd64/oocla version
v1.2.0

amd64 and arm64 builds are provided for Linux / macOS / Windows. Verify with oocla_<version>_checksums.txt.

Build from source

Requires Go 1.25 or later.

$ git clone https://github.com/kazufusa/oocla
$ cd oocla && make build   # produces bin/oocla

Usage

oocla serve [options]

Option	Default	Description
--addr	127.0.0.1:11434	Address to listen on
--claude	claude	Path to the claude executable
--bare	false	Removes everything the CLI injects, but requires API key auth (see below)

Stops on SIGINT / SIGTERM. On shutdown it waits for in-flight requests, then removes its scratch working directory.

Every turn runs claude with --no-session-persistence, so no conversation is ever written to disk. In return, a multi-turn conversation resends its whole history collapsed into a single turn, so input tokens grow with the length of the conversation.

Model names

Requested name	Value passed to claude --model
opus, sonnet, haiku, fable	Same name
Any of the above with :latest	Same name
An exact model id like claude-haiku-4-5-20251001	Passed through as is

Any tag other than :latest is a 404: there is no way to serve a pinned revision.

Endpoints

Endpoint	Description
POST /api/chat	Streaming / non-streaming. tools, format, think, options.stop
POST /api/generate	Streaming / non-streaming
GET /api/tags POST /api/show GET /api/version	Model catalog
GET /api/ps	Always empty. No model is ever resident in memory
POST /api/pull	Succeeds for known models. There is nothing to transfer
POST /v1/chat/completions	OpenAI-compatible. SSE supported
GET /v1/models GET /v1/models/{model}	OpenAI-compatible
POST /api/embeddings /api/embed /v1/embeddings	501. The claude CLI has no embedding support
POST /api/create /api/copy /api/push, DELETE /api/delete	400. No local model files exist

An empty messages array on /api/chat, like an empty prompt on /api/generate, is Ollama's preload call: it is acknowledged with done_reason: "load" ("unload" with keep_alive: 0) without invoking the model. There is nothing to actually load.

Tools

As with Ollama, the server never executes tools. It returns tool_calls; the client executes them and sends the results back as role: "tool" messages.

$ curl localhost:11434/api/chat -d '{
    "model": "haiku", "stream": false,
    "messages": [{"role": "user", "content": "What is the weather in Tokyo?"}],
    "tools": [{"type": "function", "function": {
      "name": "get_weather",
      "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}}}]
  }' | jq .message.tool_calls
[{"function":{"name":"get_weather","arguments":{"city":"Tokyo"}}}]

Environments where MCP is unavailable

Where managed settings refuse to start the MCP server, oocla detects it before the model answers and switches automatically: tool definitions go into the system prompt and --json-schema pins the answer shape (see docs/DESIGN.md for details).

Option	Description
--internal-mcp-shim-name	Register the MCP server under the name your administrator allowlisted for oocla
--internal-prompt-tools	Never use MCP; always use the prompt-based mode above (debugging)

Measured against the real models (scripts/verify-prompt-tools.sh, 2026-07):

Model	Tool call	Declining tools	Answer from result
opus	5/5	2/2	3/3
sonnet	5/5	2/2	3/3
haiku	5/5	2/2	3/3
fable	5/5	2/2	3/3

options

Ollama's options configure a local inference runner, and most of them have no counterpart in the claude CLI. Only stop is implemented; the rest are ignored with a logged warning. Nothing is rejected.

think

Value	Behaviour
absent / false	Thinking is not returned
true	Delivered in message.thinking
"low" "medium" "high"	Delivered, and the reasoning depth is set to the given level

Known limitations

By default the claude CLI injects a short fixed prompt plus a <system-reminder> carrying the logged-in email address and the date, even with an empty system prompt (about 170 tokens in total). Claude Code's full agent prompt (thousands of tokens) is successfully disabled.

To remove everything, use oocla serve --bare. In that mode the claude CLI only reads ANTHROPIC_API_KEY or an apiKeyHelper, so an OAuth login no longer works. See docs/DESIGN.md for details, including measured results for --agents / --safe-mode and friends.

Also unsupported:

• Image input. The images field is accepted but unused
• Embeddings
• Any options other than stop, e.g. num_predict temperature seed
• Continuation via context on /api/generate: it is accepted but ignored, since no session is ever kept

The final response always carries every statistics field, as clients expect from Ollama. load_duration and prompt_eval_duration are always 0: there is no model load or local prompt evaluation to time.

Development

make check   # gofmt / vet / test -race / build
make e2e     # conformance tests against a real server
make dist    # builds release artifacts into dist/

make e2e calls the real claude and costs money. It needs an authenticated environment.

A release runs when a tag starting with v is pushed. CI runs make check and then publishes the make dist artifacts unchanged. Running make dist VERSION=v1.2.0 locally produces the same files. Tags containing a hyphen, like v1.2.0-rc1, become prereleases.

Design and measured behaviour live in docs/DESIGN.md.