English | 中文

Friends: LINUX DO - 新的理想型社区

Sync pricing, channels, and models from upstream providers to your new-api instance. Supports new-api, OpenRouter, and NVIDIA NIM upstreams.

First install Bun if you don't have it.

bun install
cp config.example.yml config.yml         # edit with your config
bun sync run                             # run sync
bun sync run --only myprovider           # sync one provider
bun sync run --models "claude-*,gpt-4*"  # sync only matching models
bun sync run --verbose                   # run with debug logging
bun sync reset                           # delete all synced data

If you'd rather click than type, launch the bundled dashboard:

bun ui                          # shortcut for `bun sync ui`
bun sync ui --port 4000         # custom port (default 3000)

Then open http://localhost:3000. Everything the CLI can do is exposed through the UI, no YAML editing required:

• Dashboard: run or reset pipelines live. A streaming log panel shows every provider, model, and price in real time, with colors preserved from the CLI output. Pick which providers to touch (matches --only), restrict to specific models with glob wildcards (matches --models, e.g. claude-*, gpt-4*), and hit Start. The models filter persists per config, so switching between configs restores your last selection.
• Configuration: edit every provider, target, blacklist, price adjustment, model mapping, and per-model overrides through structured forms. Invalid YAML is validated before save and rolled back if it fails.
• Multiple configs: create named variants (debug, staging, prod) from the dropdown next to the tabs. Each is stored as config.<name>.yml next to the binary (or in the project root in dev mode) and can be switched, duplicated, or deleted without touching the filesystem directly. Cross-config settings (locale, theme, shared blacklist, shared model mapping) live in config.global.yml.
• History: browse past runs (logs/YYYY-MM-DD-*.json) with per-model pass/fail results, costs, and authenticity status. Authenticity auto-blacklist entries are manageable from the same tab.
• Themes & locales: toggle dark/light/system and switch between English and 中文; both are remembered across sessions.

The UI is bundled as a single-file binary for every platform, so the target machine does not need Bun installed. Grab the right one from dist/ after bun run build (or release artifacts) and run it directly:

./new-api-sync-linux-x64 ui      # Linux
./new-api-sync-darwin-arm64 ui   # macOS (Apple silicon)
new-api-sync-windows-x64.exe ui  # Windows

Two files are loaded on startup:

• config.yml (or a named variant config.<name>.yml): the active sync config (target, providers, etc.).
• config.global.yml (optional): cross-config settings. locale, theme, and UI state are stored here. blacklist, modelMapping and groupMapping defined here merge into every config (global blacklist entries are unioned with per-config; global mappings win on key collisions).

On upstream new-api the token is an admin access token (root for option writes). The unorouter fork accepts a scoped sync service token instead; see "Works with upstream new-api" below.

Pulls from OpenRouter. Free models (prompt=0 and completion=0) are emitted as a free tier; paid models are bucketed under a per-vendor channel and a single shared group_ratio is picked from the candidate ladder [1, 0.5, 0.25, 0.1, 0.05, 0.01] such that every kept model stays at or below canonical retail. Models that don't fit at any candidate are dropped.

Pulls from NVIDIA NIM. Text models are emitted as a free tier; image models are emitted with fixed per-request pricing (quotaType: 1) against a separate imageBaseUrl.

blacklist removes matching text models from sync. Non-text types (image, video, audio, embedding) are never filtered by the blacklist.

• Case-insensitive match against the model ID.

• Glob wildcards supported: gpt-5.*-codex, *-preview.

• Provider-scoped patterns use provider/pattern syntax. The part before the slash must match the provider's name; the part after is the glob. Example:

  blacklist:
    - nsfw # unscoped: blocks any provider's model containing "nsfw"
    - "*-preview" # unscoped: blocks any provider's preview models
    - duck/gpt-5* # scoped: only blocks gpt-5* models from the "duck" provider
    - yun/claude-*-opus # scoped: only blocks claude opus models from "yun"

A small built-in blacklist is always merged in to suppress a handful of upstream IDs that are uniformly broken or mis-typed (embedding/audio/video models served as text). You don't need to add these in your config.

groupMapping rewrites a fragment of an upstream group label before the label becomes a channel and group name (<label>-<provider>-<model>). The lane, its pricing and its upstream token are untouched; only the public name changes.

• Key is the fragment to find inside the label, matched case-insensitively. provider/fragment limits the rule to one provider.

• Value replaces just that fragment. The rest of the label survives, so duck/Antigravity: pool-a turns Gemini-CLI-Antigravity into Gemini-CLI-pool-a.

• Scoped rules run before bare ones, longer fragments before shorter; each rule runs once.

• The blacklist runs first, on the raw label and description; a rule never readmits a blacklisted group. To rename a brand lane instead of dropping it, keep the brand word out of the blacklist and add an unscoped rule for it.

• Existing channels are renamed in place: sync metadata (and the start of sync run) finds channels whose remark still carries the pre-rule label, moves their GroupRatio / UserUsableGroups / AutoGroups keys to the new name, and renames the channel row. Channel ids and stats are kept; nothing is created or deleted. Preview with sync metadata --dry-run.

• A rename whose target name already exists is skipped and logged as a collision.

  groupMapping:
    antigravity: AG # any provider: Gemini-CLI-Antigravity -> Gemini-CLI-AG
    pol/adobe: AD # only pol: image2c-adobe -> image2c-AD

priceAdjustment accepts either a single number or a keyed object:

• Number: applies uniformly. Must be in the open range (-1, 1). -0.5 = 50% cheaper, 0.1 = 10% more expensive.

• Object: keyed by model name glob, vendor name, model type, or default. Resolved in that order. Must contain a default key. Numeric keys for non-text types may go up to 1; text-type keys (vendors, model globs that resolve to text, and default) must stay below 1 so the resulting channel never costs more than calling the upstream directly. Example:

  priceAdjustment:
    default: -0.3
    image: 0.5
    anthropic: -0.1
    gpt-5*: -0.5

enabledModels accepts string globs or objects keyed by model:

enabledModels:
  - "claude-*-4-5*" # plain glob
  - model: "gpt-5"
    metadata:
      maxOutputTokens: 32768
      isReasoning: true
  - type: "image"
    model: "flux-pro"
    modelPricingGrid:
      - { "size": "1024x1024", "price": 0.04 }
      - { "size": "2048x2048", "price": 0.08 }

• metadata is forwarded to new-api's per-model metadata JSON column (consumed by client UIs to drive per-model behaviors like bumping max_tokens for reasoning models).
• modelPricingGrid defines fixed per-request pricing tables for image/video/audio models.

1. Discover: fetch models/groups from each provider, filter by vendor, blacklist, and glob patterns; splice group labels per groupMapping
2. Test: verify each model with a minimal API request
3. Build desired state: merge pricing (GroupRatio, ModelRatio, CompletionRatio), build channels and policy
4. Diff: compare desired state against current target state
5. Apply: create, update, and delete channels, models, and options
6. Cleanup: remove orphaned models

Channels are named {group}-{provider}. When a provider's models split into multiple price tiers, channels get numeric suffixes: {group}-{provider}-t0, -t1, etc. Sub-splits (caused by per-model price overrides or task model pins) add a letter: -t0a, -t0b. Priority is dynamic: cheapest groups first, faster response times get higher priority.

The sync targets upstream new-api as well as the unorouter fork. On first use it probes one fork-only route and, on a vanilla gateway, logs vanilla new-api (no sync routes) once and adapts:

Everything else (channels, vendors, option maps, ability rebuild, tokens) uses routes both have.

By default, text models whose effective ratio (group ratio x priceAdjustment) is >= 1.0 are skipped so sync never creates channels more expensive than calling the upstream directly. Non-text types (image, video, audio, embedding) are unaffected.

Disable via global config:

skipUnprofitableText: false

In addition to the unprofitable-text gate, every offer that would charge a user more than the canonical retail ratio (resolved from LiteLLM, then OpenRouter, then basellm) is dropped before sync. There is no user-facing knob for this cap; it is always 1x of canonical retail. This applies to OpenRouter paid offers.

Certain video/image models are pinned to specific channel types in new-api: sora, kling, vidu, jimeng, hailuo, seedance, veo, imagen, wan. Some also append a path suffix to the provider's baseUrl (e.g., wan becomes /alibailian). This happens automatically and produces a separate sub-channel.

During sync, model descriptions and tags are fetched from two public feeds to enrich the metadata shown in new-api:

• OpenRouter /api/v1/models: descriptions (preferred)
• basellm llm-metadata: descriptions (fallback) and tags

These calls are best-effort: failures are logged as warnings and do not block the sync. Fuzzy name matching handles version/date-suffix variants (claude-sonnet-4-5-20250929 resolves to claude-sonnet-4.5).

logs/authenticity-blacklist.json is maintained automatically by the test runner to track Anthropic Claude models from providers that failed authenticity checks. This is internal state, but entries can be reviewed and removed from the History tab in the UI.