Theoldllm Provider

OpenCode v2 provider plugin for TheOldLLM, an OpenAI-compatible API at https://theoldllm.com/v1.

The API is a plain OpenAI-compatible chat endpoint, so it works out of the box — but it publishes almost no model metadata. /v1/models returns only {id, object, owned_by, created, permission}, and it sends no rate-limit headers on any response. This plugin fills that gap: it resolves real context limits and display names from models.dev, tracks the account's 50M-token daily quota via GET /v1/usage, and models the per-model cost multiplier that decides how fast that quota burns.

Install

Add the package to plugins in opencode.json(c). OpenCode resolves and installs it itself — there is no separate npm install step.

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": ["@srah_x64/theoldllm-provider"]
}

Pin a version with @srah_x64/theoldllm-provider@0.1.0 if you want to control upgrades. Then restart the service so the provider is picked up:

opencode service restart

Registries

The same code is published to both registries under different scopes:

Registry	Package	Install note
npmjs.com	@srah_x64/theoldllm-provider	Works as-is.
GitHub Packages	@srah/theoldllm-provider	Needs an .npmrc scope line and a GitHub token with read:packages.

# ~/.npmrc — only needed for GitHub Packages
@srah:registry=https://npm.pkg.github.com
//npm.pkg.github.com/:_authToken=YOUR_GITHUB_TOKEN

Connect

Run /connect, pick TheOldLLM, and paste your key. That is the intended path: the plugin registers a theoldllm integration exposing a key method, and the provider's integrationID is what makes it appear in that list.

opencode auth login theoldllm

THEOLDLLM_API_KEY works too — it is registered as an env method, so an existing variable needs no explicit connect. Model discovery prefers the stored credential and falls back to the env var. If you set the env var it must be visible to the server process, not just your shell.

Then pick a model with /models. Discovered models appear under TheOldLLM and are referenced as theoldllm/<model-id>.

Options

{
  "plugins": [
    {
      "package": "@srah/theoldllm-provider",
      "options": {
        "baseURL": "https://theoldllm.com/v1",
        "pollMs": 300000,
        "models": ["claude-sonnet-5"]
      }
    }
  ]
}

Option	Default	Purpose
baseURL	https://theoldllm.com/v1	API endpoint.
models	discovered	Fixed model list. Skips /v1/models discovery.
pollMs	300000	How often to poll /v1/usage for quota.

What it does

Model metadata

/v1/models carries no limits, names, or modalities, so each model is looked up in the models.dev catalog — the same source OpenCode syncs from — and the real values are applied:

Field	Example
name	claude-sonnet-5 → Claude Sonnet 5 (50x)
family	claude-sonnet
time.released	2026-06-29
limit.context / .output / .input	1,000,000 / 128,000
capabilities.input / .output	text, image, pdf → text

Limits matter most. The runtime passes model.limit.context straight into the route as the true context ceiling, so a guess mis-sizes the entire context budget. Real values run from 200k (claude-haiku-4.5) to 1.05M (gpt-6-luna).

Lookup tries exact id → punctuation-folded (theoldllm writes claude-opus-4.5, models.dev claude-opus-4-5) → the base behind a known variant suffix → a dated snapshot of the same model. A variant suffix only borrows limits when the stripped base is a real catalog id, so deepseek-v4 never inherits from deepseek-v4-pro. 109 of 121 models get real limits; the other 12 are community fine-tunes and jailbreak variants that models.dev has never indexed, and those keep the 200k/32k default rather than a guess.

Cost multipliers

Every request is charged against the 50M/day quota at the model's multiplier. Measured rather than assumed: a 16-token claude-haiku-4.5 call moved daily_tokens.used by exactly 800 (16 × 50); a 49-token deepseek-v4-pro call moved it by 245 (49 × 5).

The multiplier is not exposed by the API — not in /v1/models, not in /v1/usage, not in any response header. It exists only in the pricing page's static/app.js, as a pinned table plus family defaults. multipliers.ts mirrors that logic and the model picker shows it as Claude Sonnet 5 (50x). A 0× model shows as (free).

cost is deliberately left empty. The multiplier is a quota weight, not a dollar rate, and the API never exposes a price — faking $0 would be worse than showing nothing.

Plan economics

The subscription is a flat $7/month, not a per-token rate. Combined with the multiplier this produces a real effective rate, because the daily quota is denominated in billed units: 50M billed units/day means a 5× model yields 10M real tokens/day but a 50× model only 1M.

Model	Multiplier	Real tokens/day	Real tokens/month	Effective $/Mtok
deepseek-v4-pro	5×	10M	300M	$0.023
claude-haiku-4.5	50×	1M	30M	$0.233
gpt-5.5	100×	500k	15M	$0.467
JEV	0×	—	—	free

So a 50× Claude is effectively 10× dearer per token than a 5× DeepSeek. The monthly figure assumes you exhaust the daily cap all 30 days; unused quota does not roll over, so a light month costs far more per token than this.

Variants

theoldllm ships thinking twins as separate ids (gemini-3.8-flash-thinking alongside gemini-3.8-flash). The plugin folds each twin into its base as a thinking variant, so the picker offers one row with a toggle instead of two.

A twin is only folded when the fold is free: the base must exist in the catalog and carry the same cost multiplier. That correctly leaves gemini-2.5-flash-thinking (5×) and gemini-3-flash-preview-thinking (5×) as standalone models, because their bases are 20× and folding would quietly quadruple the bill.

Effort ladders come from configuration rather than per-model probing: models.dev declares reasoning plus an effort ladder, so 43 of the 121 models pick up variants, each reflecting the levels that model accepts.

That declaration describes the upstream model, not what theoldllm honours, and the two disagree. Probing the same question plain and with reasoning_effort: "high", and comparing completion tokens:

	models	tokens
engaged	gemini-3.8-flash, gemini-3-flash-preview, claude-haiku-4.5, glm-5.2	3 → 191, 3 → 269, 5 → 39, 35 → 130
inert	claude-sonnet-5, claude-opus-5, gpt-5.5, gpt-5.6-luna, gpt-6-luna, deepseek-v4-pro	unchanged

The inert ones still declare a full ladder upstream, so left alone the catalog would hand out switches that silently do nothing — Anthropic's real control is thinking: {type, budget_tokens}, which this gateway does not map. The INERT set in variants.ts suppresses those, and a single-level ladder is dropped since it is not a choice.

Quota tracking

GET /v1/usage is the only source of quota data — the API returns no X-RateLimit-* headers on any response, so the core runtime cannot see them either.

{
  "daily_tokens": { "used": 650, "limit": 50000000, "remaining": 49999350, "reset": "midnight UTC" },
  "rate_limit": { "requests_in_window": 1, "limit": 30, "remaining": 29, "window_seconds": 60 }
}

The plugin polls it every 5 minutes (pollMs) and logs once per crossing at 80%, 95%, and 100% of the 50M daily token quota, plus an immediate re-read on any 429. Warn only — requests are never blocked. A separate 30 req/min window is reported in the same line. The daily quota resets at midnight UTC.

Streaming

Handled entirely by @opencode/ai/providers/openai-compatible, which sets stream: true and consumes the SSE frames. There is no custom stream handling here.

Development

npm ci
npm run typecheck   # tsc --noEmit
npm run build       # esbuild -> dist/index.js (single ESM bundle)
npm run verify      # smoke test: asserts the default export is a Plugin

The plugin ships as one bundled ESM file with @opencode/plugin left external, so no extension-resolution or TypeScript-at-load concerns for consumers.

To try it from a local checkout, point plugins at the folder and restart the service:

{
  "plugins": ["./theoldllm_provider"]
}

A few behaviours are worth knowing while testing:

• Model names stay short on purpose. name is what the picker renders and what its fuzzy search matches, so it carries the display label and the multiplier only — no derived price.
• A borrowed name is discarded, so gemini-2.5-flash-thinking shows as "Gemini 2.5 Flash Thinking" rather than borrowing "Gemini 2.5 Flash" and colliding with the non-thinking model. Where two ids still titleize identically (command-a vs command-a-03-2025), every member of the set falls back to its id.
• The models.dev catalog is fetched at startup and cached for 24h. A failed fetch logs a warning and reuses the previous cache rather than dropping models.

Releasing

Pushing a v* tag runs .github/workflows/publish.yml, which typechecks, builds, verifies, publishes to both registries, and attaches theoldllm-provider-<tag>.zip (built dist/, package.json, README.md, LICENSE) to a GitHub Release.

npmjs.com publishes via npm Trusted Publishing (OIDC) — there is no npm token in the workflow or in repository secrets. npm exchanges a short-lived credential scoped to this exact workflow, and generates provenance automatically. GitHub Packages uses the workflow's own GITHUB_TOKEN under packages: write, so it needs no secret either.

The two registries publish as independent parallel jobs, so a failure in one cannot silently cancel the other. The GitHub Release runs when at least one of them succeeded.

The package is published under two different scopes — @srah_x64/theoldllm-provider on npmjs and @srah/theoldllm-provider on GitHub Packages. The GitHub Packages job rewrites the name and publishConfig.registry; nothing else differs. Because GitHub Packages ships scoped npm packages as private, that job also flips the package to public after publishing.

The tag must match version in package.json; the job fails otherwise rather than publishing a mislabelled package.

npm version patch        # bumps and tags
git push --follow-tags

Trusted publishing needs the publisher configured on npmjs.com under the package's Settings → Trusted Publisher, with the organization sRaH, repository theoldllm-provider, and workflow filename publish.yml. npm CLI ≥ 11.5.1 is required (the workflow pins Node 24 and checks the npm version), and GitHub-hosted runners only — self-hosted runners are not supported.

Note that trusted publishers created after Sep 03, 2026 default to allowing npm stage publish only. For an unattended release you must also tick npm publish under Allowed actions, otherwise the publish step fails and every release needs a manual 2FA approval.

Continuous integration (.github/workflows/ci.yml) runs the same typecheck, build, verify and an npm pack --dry-run on every push and pull request.

License

MIT