omo_proxy

OpenCode (and other Anthropic-compatible clients) → Claude Code OAuth proxy with proper prompt-cache preservation.

OpenCode → omo_proxy (127.0.0.1:34156) → api.anthropic.com
                ↑
   reuses ~/.claude/.credentials.json (Claude Max subscription)

Why this exists

The Anthropic Messages API has prompt caching that can give a ~9x cost reduction on long contexts (cache_read is 0.1x the input price). A naive proxy that rewrites or shuffles the request body (billing headers, session ids, cache_control markers) silently destroys cache prefixes and forces every request to be billed at full price. On a Claude Max plan that drains the rate limit in hours.

This proxy specifically preserves prompt caching while still authenticating against Claude Code's first-party OAuth token.

Features

• Stable cache prefix. Billing/session strings persisted in proxy-state.json survive process restarts, so previously written cache entries can still be reused.
• cache_control cap = 4 (Anthropic max), not 1. When trimming is needed, keeps the tools and system anchors plus the latest messages markers.
• Stable metadata.user_id. Caller-supplied volatile fields are dropped.
• OAuth token persistence. Refreshed tokens are written back to ~/.claude/.credentials.json, so a proxy restart never invalidates the refresh chain.
• Tool name normalization. OpenCode lowercase tools are mapped to Claude Code's CapitalCase set; MCP tools (mcp__*) pass through.
• Usage logging. Every request prints a [usage] line including cache_read, cache_create, cache_create_5m, cache_create_1h, total_in, etc.
• Hardened launcher. ensure-claude-proxy.ps1 checks the exact proxy.js is the listener, not just "something on port 34156".

Verified cache behavior

Log excerpt from a real session (224k token system+tools prefix):

#1: in=6  total_in=224281  out=1060  cache_read=0       cache_create=224275
#2: in=6  total_in=225387  out=467   cache_read=224275  cache_create=1106
#3: in=1  total_in=226023  out=517   cache_read=225381  cache_create=641

→ From request #2 onward Anthropic returns the entire prefix as cache_read, billing at 0.1x. Cost on the same conversation drops by ~9x compared to the broken-cache baseline.

Layout

omo_proxy/
├── proxy.js                    main Node.js proxy (804 lines)
├── start-proxy.cmd             foreground launcher (Windows)
├── opencode-cc.cmd             one-shot env wrapper (cmd)
├── opencode-cc.ps1             one-shot env wrapper (PowerShell)
├── shims/
│   ├── opencode.cmd            transparent shim for cmd / PowerShell
│   ├── opencode                transparent shim for Git Bash / WSL
│   └── ensure-claude-proxy.ps1 helper that ensures the right proxy.js is running
└── go/                         legacy Go implementation (not maintained)

Requirements

• Node.js 18+ (uses built-in http, https, crypto, fs).
• Claude Code installed and logged in (~/.claude/.credentials.json must exist with a valid claudeAiOauth block).
• A Claude subscription that includes the model you select in OpenCode (e.g. Max for Opus).

Quick start (Windows)

git clone https://github.com/winglock/omo_proxy "$env:USERPROFILE\omo_proxy"

# (option A) just run the proxy in a foreground window
cd "$env:USERPROFILE\omo_proxy"
node proxy.js

# (option B) install the transparent opencode shim
Copy-Item "$env:USERPROFILE\omo_proxy\shims\opencode.cmd" "$env:USERPROFILE\bin\"
Copy-Item "$env:USERPROFILE\omo_proxy\shims\ensure-claude-proxy.ps1" "$env:USERPROFILE\bin\"
# make sure $env:USERPROFILE\bin is ahead of npm's path so 'opencode' resolves to the shim

If you keep the repo somewhere other than ~/omo_proxy, set:

setx OMO_PROXY_SCRIPT "C:\path\to\omo_proxy\proxy.js"

Quick start (macOS/Linux)

The Node.js proxy works on any platform. macOS users get the bonus that Claude Code stores credentials in Keychain too — currently only the Go variant reads Keychain; on macOS you'll either symlink the credentials JSON into ~/.claude/.credentials.json or extend proxy.js's TokenManager to call security find-generic-password.

git clone https://github.com/winglock/omo_proxy ~/omo_proxy
cd ~/omo_proxy
node proxy.js &

ANTHROPIC_BASE_URL=http://127.0.0.1:34156/v1 \
ANTHROPIC_API_KEY=proxy-placeholder \
opencode

OpenCode config snippet

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "anthropic": {
      "options": {
        "baseURL": "http://127.0.0.1:34156/v1",
        "apiKey": "proxy-placeholder"
      }
    }
  }
}

Verifying the cache is alive

After a few requests:

Select-String -Path ".\proxy.log" -Pattern "\[usage\]" | Select-Object -Last 10

Healthy pattern: cache_read is roughly equal to total_in from the second request onward, and cache_create is small.

If cache_read=0 on every request, something downstream of this proxy is mutating the prompt prefix per call. Check your client first; this proxy stabilizes the parts it controls.

Tool filtering

Anthropic's classifier rejects unknown tool names (treats them as third-party). This proxy keeps only Claude Code's canonical set plus MCP passthrough:

kept: Bash, Read, Glob, Grep, Edit, Write, Task, TodoWrite, WebFetch, WebSearch, Skill, BashOutput, NotebookEdit, mcp__*

dropped: LSP tools (lsp_*), ast_grep_*, session_*, background_cancel, look_at, skill_mcp, websearch_web_search_exa, grep_app_searchGitHub, context7_*, etc. The dropped tools simply aren't available to the Claude model — non-Anthropic providers in the same OpenCode session keep all of them.

Endpoints

Path	Behavior
GET /health	Local-only check. Returns 200 {"status":"ok"} if the loaded token is more than 5 minutes from expiry, 503 {"status":"token_refresh_required"} otherwise. Does not call upstream.
POST /v1/messages*	Transformed and forwarded. Streaming SSE supported.
anything else under /v1/	Auth-rewritten and forwarded as-is.

Files this proxy reads/writes

Path	Direction	Contents
~/.claude/.credentials.json	read + write	Claude Code OAuth tokens. Updated on refresh.
./proxy-state.json	created on first run	Persisted billing lines + session id used to keep cache prefix byte-identical across restarts. Do not commit.
./proxy.log	append	Request/usage log; rotated at 10 MB.

Disclaimer

This is a personal-use local proxy for an individual subscriber's own first-party OAuth token. It is not affiliated with Anthropic or OpenCode. Don't use it to share credentials, redistribute access, or anything that violates the Claude Code TOS.

License

MIT