English · 简体中文

Log in to any model gateway with a base URL and an API key — every chat model it serves becomes selectable.

           

An omp (@oh-my-pi/pi-coding-agent, aka oh-my-pi) extension that turns any "base URL + API key" gateway into a first-class login provider — CLIProxyAPI, LiteLLM, one-api, new-api, the Volcengine Ark coding plan, in-house routers — as long as the endpoint answers GET <base URL>/v1/models and speaks Anthropic Messages (or OpenAI Chat Completions).

One installation fronts as many gateways as you configure. You name each one during login, and that name is the model prefix.

Install

Needs omp ≥ 18.2.3 (@oh-my-pi/pi-coding-agent), the agent that loads this extension.

omp plugin install omp-model-gateway

Also installable straight from git: omp plugin install github:magic3007/omp-model-gateway. Do not shorten the npm name — the package omp-gateway is an unrelated project.

Coming from the earlier omp-gateway id? Move once — omp plugin uninstall omp-gateway, then install as above. Nothing else changes: the config file and its gateway entries stay where they are.

Log in

omp login                            # pick "Custom gateway (name + base URL + API key)"
                                     #   ① name        e.g. litellm
                                     #   ② base URL    e.g. https://gw.example.com
                                     #   ③ key         LITELLM_API_KEY, or sk-…, or !pass show gw
                                     #   ④ extra ids   Enter for none
# → Saved gateway "litellm" (48 chat models) … select them as litellm/<id>. Key stored as the reference LITELLM_API_KEY.

The name, URL and key are validated together with one model-list call before anything is stored, so a wrong key or an unreachable URL fails the login instead of adding a broken gateway. Repeat the login for each gateway you want — all of them are registered on every start. What lands in the config file is the reference you typed, not the secret (Key references). (omp's own line about credentials being stored, or not, refers to its credential store; this plugin keeps the reference in its own file and lets omp resolve it.)

Enable the models

If your ~/.omp/agent/config.yml already has an enabledModels allow-list, add one glob per gateway name — without that list, every discovered model is already selectable:

enabledModels:
  - litellm/*
  - volcengine/*

modelRoles:                             # optional aliases
  gw-opus: litellm/claude-opus-4-1:auto
  ark-code: volcengine/ark-code-latest:auto   # :auto is omp's thinking-level suffix

Then use them like any other model:

omp models list litellm
omp -p --model litellm/claude-opus-4-1 "reply with exactly: OK"

What you get

Item	Value
Login entry	/login (TUI) or omp login (terminal) → Custom gateway (name + base URL + API key)
Login input	name, base URL, a key or a key reference (environment variable name, !command, literal), optional extra model ids
Model prefix	<name>/<model id>, e.g. litellm/claude-opus-4-1, volcengine/ark-code-latest
Wire	anthropic-messages by default; set openai-completions per gateway in the config file
Model source	GET <base URL>/v1/models, re-fetched on login and on discovery refresh
Gateways per install	unlimited — one entry per name in the config file

Discovery drops what is not a usable chat model (image/video/speech ids, entries the gateway marks as shut down or deprecated) and uses the context window and output cap the gateway reports, falling back to 200k / 64k.

Model ids — and so selectors — are exactly what the gateway lists. The name omp shows (status line, selector) is the gateway's display name when it reports one; otherwise the id without its upstream-source prefix (devin/claude-opus-5-5 shows as claude-opus-5-5). When two ids share that short name (devin/gpt-6-luna beside gpt-6-luna), the prefixed one names its source: gpt-6-luna (devin).

Managing gateways

Everything lives in one JSON file, ~/.omp/agent/gateways-login.json (0600):

{
  "providers": [
    { "name": "litellm", "baseUrl": "https://gw.example.com", "api": "anthropic-messages",
      "apiKey": "LITELLM_API_KEY", "extraModels": [] },
    { "name": "volcengine", "baseUrl": "https://ark.cn-beijing.volces.com/api/coding",
      "api": "anthropic-messages", "apiKey": "VE_CODE_API_KEY",
      "extraModels": ["ark-code-latest", "doubao-seed-2.0-pro", "glm-5.3"] }
  ]
}

• Add — run the login entry again with a new name.
• Update — run it with an existing name: Enter keeps the stored URL and key reference, - clears the extra ids.
• Remove — delete its object from the file and restart omp.
• extraModels — model ids the gateway serves but does not list (plan-only aliases, e.g. the Ark coding plan's ark-code-latest).
• api — set openai-completions for gateways that only speak OpenAI Chat Completions.

Key references

apiKey is a reference, not necessarily a key: omp resolves it every time the key is used, with the same rules it applies to models.yml (resolveConfigValue).

You write	omp sends	Use it for
LITELLM_API_KEY	that environment variable	the usual case — the secret stays in your shell environment
!gpg -q -d ~/.omp/keys/gateway.gpg	the command's stdout, trimmed	secrets in an encrypted file, pass show …, op read …, a keychain lookup
sk-…	the text itself	quick trials; the file then holds the secret

So the file is safe to track in a dotfiles repository: it names where the key comes from, and never has to contain it. The login entry writes a reference too — paste a key that some environment variable already holds and it stores that variable's name instead, saying so in its output. See SECURITY.md for the full posture.

Two things to know about the environment form: the variable must be exported in every shell that starts omp, and a name nothing sets is sent as the literal key (omp's rule), which the login flow reports as such. A !command runs as your user through /bin/sh -c; omp caches its output per process, so an interactive gpg or op prompt happens once per omp run — the login itself runs it once more.

Hand-editing works: extraModels and api are the fields the login flow does not ask about. Every row is validated on its own, so a mistake costs that gateway only — but the login entry refuses to run while any row is unusable, rather than writing the file back without it.

Base URLs are stored as the gateway root (https://gw.example.com, a trailing /v1 is stripped); the plugin adds /v1/models for discovery, and the /v1 the OpenAI wire needs. After hand-editing, run omp models refresh so the cached catalog is re-read.

Development

git clone https://github.com/magic3007/omp-model-gateway
cd omp-model-gateway
bun install              # dev only: TypeScript and the omp host types
omp plugin link .        # edits in src/ take effect on the next omp start
bun run typecheck && bun test

No runtime dependencies: the entry point is TypeScript source (package.json#omp.extensions: ["src/index.ts"]) loaded directly by omp's Bun runtime. Host types come through import type only.

Remove: omp plugin disable omp-model-gateway (keep, stop loading) or omp plugin uninstall omp-model-gateway.

See CHANGELOG.md for release notes, CONTRIBUTING.md before opening a pull request, and SECURITY.md for how API keys are stored.

License

MIT