Claude Code Ollama Proxy

A local HTTP server that speaks the Ollama API and drives the Claude Code CLI, so Visual Studio 2026's GitHub Copilot Chat “Bring Your Own Key” (BYOK) Ollama provider can use Claude Code — including Copilot's native in-chat diff / Accept-Reject / “Accept all” / rollback UI — without installing any extra Visual Studio extension.

Unlike proxies that forward to cloud AI providers, this one launches the local claude CLI (via the AJGit.Claude.AgentSdk SDK) and nothing else, so it uses your existing Claude Code login.

How it works

Copilot's BYOK Ollama provider splits discovery from chat:

• Discovery (Ollama-native): GET /api/version, GET /api/tags, POST /api/show.
• Chat: POST /v1/chat/completions — OpenAI shape, streamed as SSE.

The proxy detects the Copilot mode from the tools on each request:

Mode	How it's detected	What Claude does
Ask	No edit/terminal tools	Read-only chat (Claude plan permission mode)
Plan	manage_todo_list etc. but no edit tools	Read-only research
Agent	Edit tools present (apply_patch, …)	Full work via the tool-call bridge (below)

The tool-call bridge (Agent mode)

Copilot's native diff/accept UI only lights up when the model emits tool_calls to Copilot's edit tools, which Copilot applies via WorkspaceEdit. So in Agent mode the proxy:

1. Registers the tools Copilot sent as in-process MCP tools on a live Claude session (mcp__copilot__<tool>), and disables Claude's own file-mutating tools so edits must flow through them.
2. When Claude calls one, the MCP handler emits an OpenAI tool_call to Copilot, ends the response with finish_reason: "tool_calls", and blocks — keeping the Claude session alive in memory.
3. Copilot applies the edit (its native diff/accept UI), then sends a follow-up request with the role:"tool" result. The proxy correlates it by tool_call_id, unblocks the handler, and the same Claude session continues — streaming more tool calls or the final answer.

Claude's own read-only tools (Read/Grep/Glob/LS) stay enabled for efficiency.

Prerequisites

• .NET 10 SDK.
• Claude Code CLI installed and on PATH, and authenticated:
  • npm install -g @anthropic-ai/claude-code
  • claude login (or set ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN in the proxy's environment).
• A Visual Studio 2026 / VS Code build whose Copilot Chat supports the Ollama BYOK provider (and Agent mode for it).

Run

dotnet run --project ClaudeCodeOllamaProxy

The proxy listens on http://127.0.0.1:11434 (the default Ollama port). Override with ASPNETCORE_URLS if needed.

Configure Visual Studio 2026

1. Copilot Chat → Manage Models → add an Ollama provider.
2. Set the endpoint to http://localhost:11434.
3. The claude-* models appear; pick one and chat. Switch to Agent mode to let Claude edit files — changes show up in Copilot's normal changed-files / diff / accept-reject UI.

Effort levels

The claude CLI supports an effort level per session (--effort low|medium|high|xhigh|max), but Copilot's Ollama provider has no effort control — so the proxy exposes each model at each supported effort as a separate selectable model. In Copilot's model picker you'll see entries like:

Opus 4.8 (1M context) · high effort      (id: opus:high)
Sonnet 4.6 (1M context) · xhigh effort   (id: sonnet:xhigh)

Pick the one you want; the proxy parses the :<effort> suffix and passes --effort to the CLI. /api/show reports each model's effort, full supported_efforts list, display_name, and context size. Requests for a bare model id (no suffix) use Claude.DefaultEffort.

Configure under Claude in appsettings.json:

"DefaultEffort": "high",
"SupportedEfforts": [ "low", "medium", "high", "xhigh", "max" ]

SupportedEfforts controls which variants are listed. Only the CLI's documented flag values are valid — ultracode and auto are interactive-/effort-only and are dropped (the CLI ignores unknown --effort values and falls back to its default).

Configuration (appsettings.json)

Models are retrieved from the claude CLI at startup (not hardcoded). appsettings.json only holds behavior overrides and capability defaults:

• Claude.WorkingDirectory — explicit working directory. When empty, the proxy infers the solution directory from absolute paths in the request (walking up to the nearest *.sln/.git). If your edits land in the wrong place, set this to your solution folder.
• Claude.CliPath — path to claude if not on PATH.
• Claude.MaxTurns, Claude.BridgeTerminalTool (route run_in_terminal through Copilot), Claude.ModelsCacheTtlMinutes.
• Claude.McpToolTimeoutMs — sets the CLI's MCP_TOOL_TIMEOUT (wall-clock cap on a bridged Copilot tool call, so long build/test/debug tools don't time out; default 600000 = 10 min; 0 = CLI default ~28h).
• Claude.MaxMcpOutputTokens — sets the CLI's MAX_MCP_OUTPUT_TOKENS so large read_file/search results aren't truncated at the ~10k-token default (default 100000; 0 = CLI default).
• ModelDefaults — per-model DisplayName (friendly version label, e.g. "Opus 4.8"), ContextLength, and Capabilities, keyed by case-insensitive substring of the model id ("*" is the fallback). The CLI doesn't report these, so they populate /api/show, /api/tags, and the friendly display labels (the "tools" capability enables Agent mode; "vision" enables image input).

Logging

Logs go to the console and to rolling files under Logs/ via Karambolo.Extensions.Logging.File. The listening URL is printed to the console on startup by default.

• File naming: a unique file per launch (<launch> token), rolling daily (<date>) and at 10 MB (<counter>): ClaudeCodeOllamaProxy-<date>-<launch>-<counter>.log. Newest 3 files are kept (swept at startup; Karambolo has no native retention).
• By default, high-level events log at Information (model, detected mode, resolved working directory, effective Claude options, bridged tool calls, exceptions + how they were handled).

Most-verbose logging (CLI command line, stdout/stderr, HTTP requests/responses)

Set the minimum log level to Trace. The simplest switch (keeps ASP.NET/Kestrel quiet):

# environment variable
Logging__LogLevel__Default=Trace dotnet run --project ClaudeCodeOllamaProxy

or in appsettings.json set "Logging": { "LogLevel": { "Default": "Trace" } }.

Levels:

• Information (default): per-request model/mode/effort/working-dir, effective Claude options, each bridged tool_call name + id, tool-result sizes, client disconnects (HTTP 499), exceptions.
• Debug: the system prompt and prompt sent to Claude, bridged tool-call arguments, and tool-result content (all truncated). Set Logging__LogLevel__Default=Debug for these.
• Trace (everything):
  • HTTP requests and responses (incl. streamed SSE/NDJSON chunks) — category ClaudeCodeOllamaProxy.Wire.
  • The claude CLI command line (Starting CLI: …/claude.exe --output-format stream-json …) and CLI stdout — category Claude.AgentSdk (its transport logs the spawn command at Debug).
  • CLI stdin/stdout/stderr as the proxy sees them — category ClaudeCodeOllamaProxy.ClaudeCli. (One-shot Ask/Plan requests use --print, so there is no stdin stream; the bridge session does.)

To raise just one area instead of everything, set that category, e.g. Logging__LogLevel__ClaudeCodeOllamaProxy=Trace (proxy wire + CLI I/O + options) or Logging__LogLevel__Claude.AgentSdk=Debug (CLI command line). Microsoft.AspNetCore stays at Warning regardless, so Kestrel noise is suppressed.

Image input

Best-effort: images Copilot sends (data: URLs / base64) are written to temp files and referenced by path so Claude's Read tool can view them; http(s) image URLs are passed through as references.

Troubleshooting

• debugger_launch_unit_test (and similar) "caps execution at 30 seconds": that cap is in Copilot's own tool (VS-side), not this proxy or the claude CLI — the proxy can't extend it. Claude.McpToolTimeoutMs only governs how long the proxy/CLI will wait for a bridged tool result (already generous); it does not change Copilot's per-tool caps.

Limitations

• In-flight bridge state is in memory only. Restarting the proxy mid-tool-loop drops the active Claude session; the next Copilot request falls back to starting a fresh turn.
• The Claude CLI may impose its own timeout on a long-pending tool call (the SDK imposes none).
• Image input is best-effort (no native multimodal input in the CLI SDK).
• Streaming is message-granular (per assistant message), not token-by-token.
• Requires a Copilot build where the Ollama BYOK provider (and its Agent mode) is enabled.

References

• SDK used: AJGit.Claude.AgentSdk · NuGet
• File logging: Karambolo.Extensions.Logging.File
• Other C# Claude-CLI SDKs considered: 0xeb/claude-agent-sdk-dotnet, zxyao145/claude-code-sdk-csharp, managedcode/ClaudeCodeSharpSDK, gunpal5/claude-agent-sdk-dotnet
• Official Agent SDK (Python/TS only): https://code.claude.com/docs/en/agent-sdk
• Reference projects: vs2026-copilot-deepseek-v4 (proxy structure), zmy15/DeepSeek-v4-for-VisualStudio (tool-loop patterns), dliedke/ClaudeCodeExtension (the extension this obviates)
• Copilot / Ollama protocol: microsoft/vscode-copilot-chat, VS Code BYOK docs, Ollama API