corti-bridge

Run Claude Code on Corti Models — a local gateway that translates Anthropic Messages to OpenAI Chat Completions and back, including streaming. No dependencies, no build step.

 

What it does

• Bidirectional translation — Anthropic Messages ⇄ OpenAI Chat Completions both ways: system prompts, tools, model names, tool_use/tool_result pairing, thinking config, and images
• Streaming — SSE; parallel tool calls round-trip by index
• Auth swap — discards the client's token, sends CORTI_BEARER upstream; writes nothing to any settings.json, so a plain claude session stays on Anthropic models
• Model tiers — maps fable/opus/sonnet/haiku to Corti S1 models by model-ID shape, so a new generation needs no code change
• WebSearch — converted to a function tool, results intercepted via Tavily with a keyless DuckDuckGo fallback
• Resilient gateway — keeps running between sessions, auto-restarts when stale, and retries transient upstream 5xx bursts before the client ever sees them
• Diagnostics & model picker — corti-bridge doctor checks the install, gateway, and state; corti-bridge models picks which Corti S1 model backs each tier

 

Quick start

Requires a Corti account (CORTI_BEARER/CORTI_BASE_URL from npx @corti/cli models init — see the Corti Models docs) and Node.js 20+. POSIX sh, so it runs under any shell.

git clone https://github.com/corticph/corti-bridge
cd corti-bridge
./setup.sh     # checks deps + creds, installs the wrapper, picks models
corti-bridge   # starts the gateway if needed, then launches claude

Re-running ./setup.sh is safe; --yes skips the prompts for an unattended install.

 

Usage

corti-bridge                 # start the gateway if needed, then launch claude
corti-bridge models          # pick which Corti model backs each tier
corti-bridge doctor          # diagnose the install, gateway, and state
corti-bridge theme           # print the lime-mascot TUI theme + install steps
corti-bridge restart         # stop then start (needs CORTI_BEARER/CORTI_BASE_URL)
corti-bridge --stop          # stop the gateway
corti-bridge help            # full reference
corti-bridge --anthropic     # pass-through mode (escape hatch; some features unavailable)

The gateway runs on 127.0.0.1:4192 and outlives any single session; the wrapper auto-restarts it when stale (moved base URL, changed source).

 

Staying up to date

corti-bridge checks once a day whether main has moved, and tells you at the end of a session:

corti-bridge: 3 new commits on main.
  cd ~/corti-bridge && git pull && ./setup.sh

There is no restart step — pulling changes the gateway's source, and the next launch picks it up automatically. corti-bridge doctor reports the same thing on demand.

The check is a git fetch against the clone's own origin — no account, no telemetry, nothing sent anywhere. Set CORTI_NO_UPDATE_CHECK=1 to turn it off; it is already off on any branch but main.

 

How it works

The default openai mode is a translating gateway: Anthropic Messages in, OpenAI Chat Completions to CORTI_BASE_URL, responses translated back. --anthropic is a thin pass-through with no translation — an escape hatch only, since streaming drops input-token accounting (context and cost readouts stop working). See GUIDE.md for the full mechanics, trade-offs, and known degradations.

Model mapping lives in ~/.corti-bridge/models.env, written by setup.sh from Corti's catalog. The wrapper exports it as process-scoped env, so Corti model IDs never leak into a plain claude session.

 

Files

corti-bridge/
├── gateway.mjs         # The server: routing, phases, upstream client, logging
├── translate.mjs       # All wire-format logic: request/response/SSE translation, errors
├── bin/corti-bridge    # Wrapper: starts the proxy, launches claude
├── setup.sh            # Installer
├── lib/                # Shell + JS helpers sourced by setup.sh and the wrapper
└── test/               # smoke, translate, retry, dispatch, models — each a self-contained suite

After install, the runtime layout is:

~/.corti-bridge/             # Proxy state (models.env, profile.env, gateway.log)
~/.local/bin/corti-bridge    # The wrapper

 

Environment

Read directly from the shell — no local secrets file.

Var	Required	Notes
CORTI_BEARER	yes	Sent upstream; the client's own token is never used
CORTI_BASE_URL	yes	Must match https://ai.<env>.corti.app/v1
CORTI_PORT	no	Proxy bind port, default 4192
CORTI_HOST	no	Proxy bind address, default 127.0.0.1
CORTI_DEBUG	no	Any value except 0/false/no/off enables request/response logging
CORTI_ADVISOR	no	auto (default) — the consult_advisor advisor tool is on in openai mode, off in anthropic mode; on/off force both modes. Spawns a headless Opus-tier consult on the request path. See GUIDE.md for the advisor mechanism, the experimental anthropic-mode caveat, and CORTI_ADVISOR_* tuning.

ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, CLAUDE_CODE_ATTRIBUTION_HEADER=0, and CLAUDE_CODE_DISABLE_1M_CONTEXT=1 are exported by the wrapper itself — plumbing this tool owns, not something you configure. The full variable reference (search backends, reasoning mode, debug caps, state directories) is in GUIDE.md.

 

Upgrading

1. git pull
2. ./setup.sh — refreshes the wrapper, reports what's already current, re-asks nothing you've already answered
3. corti-bridge — the wrapper auto-restarts a stale gateway
4. ./setup.sh --fresh if Corti has changed what it serves since you last ran it

./setup.sh --uninstall removes the wrapper and the PATH block it added. It leaves ~/.corti-bridge alone (your model mapping and profile choice) and prints the path so you can delete it yourself.

 

Reference

GUIDE.md covers the rest: the translation surface (system-prompt merge, tool mapping, thinking config, images), the model-ranking algorithm and capability derivation, the fingerprint probe that de-aliases models, debug-logging internals and sensitivity, the profile choice, the full environment reference, and the known degradations in openai and anthropic mode.