GWDG's SAIA API retired four models in a single wave this October — apertus-70b, devstral-2-123b, gpt-oss-120b and mistral-medium-3.5 vanished from GET /v1/models between one Friday and the next. A coding agent configured against any of them stops dead: every request returns a model error, and the failure looks like your bug, not theirs. Meanwhile the service hands out million-token-context models for free, hosted on German public-sector infrastructure.
This is the maintenance problem the SAIA plugin family solves. Three open-source plugins — one per coding agent — keep pi, OpenCode and zot pointed at the live catalog of SAIA, the OpenAI-compatible API layer that GWDG (Gesellschaft für wissenschaftliche Datenverarbeitung, Göttingen) operates on the Academic Cloud. They take three different architectural bets on the same question, and the contrast is the interesting part.
SAIA is the API behind Chat AI, KISSKI's LLM service for German academia. The properties that matter for coding agents:
/v1 endpoints — /chat/completions, /completions, /embeddings, /models. Anything that speaks the OpenAI chat-completions protocol can use it. It does not implement the Responses API, so agents needing that API surface are out of luck.$ curl -is -H "Authorization: Bearer $SAIA_API_KEY" \
https://chat-ai.academiccloud.de/v1/models | grep -i ratelimit
x-ratelimit-limit-minute: 10
x-ratelimit-limit-hour: 200
Ten requests per minute is tight for an agentic loop with sub-minute tool round-trips. The plugins therefore treat the quota as a first-class concern: pi ships usage tracking and a skills-driven health check, OpenCode gets a TUI quota widget fed by a response-observing local server, zot carries the limits in its typed catalog so the agent can pace itself before a 429 arrives.
Access is an Academic Cloud account plus a free API key requested through the KISSKI LLM service form — approved within days, time-limited, renewable.
| pi-saia-plugin | opencode-saia-plugin | zot-saia-plugin | |
|---|---|---|---|
| Language | TypeScript (bun) | TypeScript (bun) | Go |
| Integration | generated models.json provider + extensions | plugin API runtime wrapper | compiled agent + static catalog |
| Catalog strategy | curated facts layer, generator writes provider config | runtime derivation from /v1/models | typed Go structs projected to JSON |
| Extra surfaces | skills, aliases, reasoning-level maps, memory | quota TUI, transports, settings | facts JSON, wire-level tests |
| Update trigger | drift check + sync script | live, on discovery | drift check + rebuild |
pi is the richest: an install task runs the generator, which reads the curated catalog and writes the provider block into pi's config, then ships skills that know how to re-check drift, refresh aliases and report quota state. The catalog carries judgment values — cost estimates per model class, reasoning-level maps, developer-role compatibility flags — that the API itself never exposes.
OpenCode went the opposite direction after a ground-up rebuild: its runtime derives model metadata live from the API response and hard-codes almost nothing. Install a static file set, point the wrapper at the endpoint, and discovery does the rest.
zot, being Go, wants types: every model is a struct, the catalog compiles, and a facts JSON projection mirrors it for tooling. The binary embeds the catalog, so staleness is a compile-time problem.
Since mid-2026 the live SAIA catalog has settled at roughly fourteen models — but the membership churns. glm-4.7, a 2.4-trillion-parameter Qwen, medgemma and several 122b-class variants disappeared over summer; the October wave cleared four more. DeepSeek V4 Flash (1M context, fast, free) is currently the sensible default; Mistral Large 4 at 1 048 576 tokens is the flagship reasoning pick.
Three things bite you when a catalog moves this fast:
glm-4.7 fails on first use. Aliases (best-for-reasoning, budget, fastest) must be re-pointed, never left dangling.The defence is boring and effective: a collector script pulls the live catalog, a sync script merges it into the curated layer, a drift checker diffs curated against live — all runnable offline against a fixture via SAIA_MODELS_DATA, so tests never depend on network or rate limits.
The family embodies two defensible answers to "where should model metadata live?"
Curated facts (pi, zot) means someone vouches for each value: cost estimates, latency class, reasoning-level support, aliases. It survives API outages, works offline, and can express judgement the API doesn't carry — but it will go stale, so it needs the drift machinery and someone to run it.
Runtime derivation (opencode) means the API is the source of truth: names, attachment support and options arrive with the catalog, nothing to rot. The cost is that you only ever know what the API tells you — no cost estimates, no aliases with intent, and a surprise model appears with no metadata until you look.
Neither is wrong. The interesting engineering constraint is that both live in the same organisation and must not silently disagree: the pi and zot catalogs are reconciled against the same live fetch, and any divergence between "facts" and "derivation" surfaces as a test failure, not a user bug report. That reconciliation discipline — facts go to the shared remote, taste stays with the maintainer — is what keeps three independent codebases telling users the same story about one endpoint.
Curated catalogs have a second failure mode: the machines. The fleet runs these plugins across workstations and web hosts via Ansible, and the obvious deployment is also the broken one — a package-install task guarded by "is it already present?" will happily skip every update forever, because presence is not currency. The fix is a small, opt-in refresh step that re-runs the installer for branch-tracking packages, plus a verification loop that asks the host (not the playbook) which commit it actually runs:
git -C ~/.pi/agent/git/github.com/tobias-weiss-ai-xr/pi-saia-plugin log --oneline -1
# 1a0bc31 chore: finish WIP — refresh legacy surfaces to the live catalog
Catalog refreshes then become routine: sync, test, push, deploy, verify — the same shape the OpenCode side follows with its file-set rsync and its own update script pinning the branch.
The whole family runs on a free SAIA key: an Academic Cloud account (federated login with your university credentials, or a self-registered Academic ID) plus the request form on the KISSKI LLM service page — choose API access to our chat service, use the same e-mail address as your Academic Cloud account, and the key arrives by mail within a few days. Keys are time-limited; the mail includes the renewal procedure. Model list and limits are one authenticated curl away, and every plugin's onboarding path starts exactly there.
The three repos share a maintenance routine rather than shared code — no common library, deliberately, because the agents' integration models are too different. What is worth stealing is the pattern: exact token counts over round numbers, tests that parse your own documentation, offline-capable drift checks, and deploys that verify what the host actually runs. Catalog churn is not an exception in the academic-AI world; it is the steady state, and the tooling that assumes churn is the tooling that survives it.