Skip to content

API access

Every account on mit.nonlocally.org can call the platform from code with a personal API key, through the same OpenAI-compatible surface the chat UI uses. Each claim on this page carries the date it was checked against the running platform; tests/test_goal_20260904.py pins the page to the committed configuration so it cannot silently drift.

1. Create a key

Settings → Account → API keys → Create new secret key. Copy the key (sk-…); the same pane lets you regenerate or delete it later.

Measured 2026-09-04

API keys are enabled for every user (auth.enable_api_keys = true, user permission features.api_keys = true) and no endpoint restrictions are configured, so a key reaches every endpoint your account can. Browser sessions expire after 4 weeks (auth.jwt_expiry); keys do not carry an expiry — revoke by deleting or regenerating.

2. Base URL, endpoints, auth
Base URL https://mit.nonlocally.org/api
Auth header Authorization: Bearer sk-…
List models GET /api/models — exactly the models your account sees in the picker
Chat POST /api/chat/completions — OpenAI chat-completions shape, stream: true or false 
export OWUI_KEY="sk-…"
curl -sS https://mit.nonlocally.org/api/chat/completions \
  -H "Authorization: Bearer $OWUI_KEY" -H "Content-Type: application/json" \
  -d '{"model":"claude-haiku-4-5","messages":[{"role":"user","content":"Reply with the single word OK"}],"max_tokens":5}'

from openai import OpenAI
client = OpenAI(base_url="https://mit.nonlocally.org/api", api_key="sk-…")
print([m.id for m in client.models.list().data][:5])
r = client.chat.completions.create(model="claude-haiku-4-5",
                                   messages=[{"role": "user", "content": "Reply with the single word OK"}],
                                   max_tokens=5)
print(r.choices[0].message.content, r.usage)

Measured 2026-09-04

With a user key: GET /api/models → 200 (47 models for an admin account — visibility is per account); the request above on claude-haiku-4-5 → OK with a usage block (13 prompt / 4 completion tokens). An invalid key → 401. The model gateway behind the API (LiteLLM) is reachable only inside the cluster; this /api surface is the one public API.

Model ids are the ones the picker shows: plain routes such as claude-sonnet-5, gpt-5, grok-4.6, glm, [local] Qwen3 235B, and the presets (code-agent, formal-verifier, photonics-designer, …). Every call — chat or API — draws down the same subscription allowance: see Usage & metering.

3. Function calling: which models honour a tools parameter

If your code passes OpenAI-style tools, the call is only useful on a model that returns tool_calls. The list below is the set of routes whose FORCED-tool_choice probe returned a real tool call, run from inside the cluster against the gateway (record: tests/data/tool_calling_probe.json, measured 2026-09-01T23:29Z, glm folded in 2026-09-03 via #506; the test above fails if this list and the record diverge).

claude-opus-5 · claude-sonnet-5 · claude-haiku-4-5 · claude-opus-4-6 · claude-sonnet-4-6 · claude-opus-4-5-20251101 · claude-sonnet-4-5-20250929 · gpt-5 · o3 · or-gpt-5 · or-gpt-5-mini · or-o3 · or-o4-mini · grok-4.6 · qwen3-5-397b · [hf] Qwen3.5 397B · workspace-agent · glm

Routes added after that probe and declared tool-capable at the gateway but not yet in the committed record: [local] Qwen3.8 27B (FORCED probe through the proxy 2026-09-03), glimmer. Reasoning-only and prover routes (mathstral, deepseek-prover-7b, deepseek-r1, [local] Qwen3 235B) answer in prose and should be called without tools.

Server-side tools run in the chat UI, not through a bare API call

The presets' built-in tools — marimo notebooks, GDSFactory, VVUQ Lean compile, GitHub — execute inside Open WebUI's chat pipeline. A plain POST /api/chat/completions to code-agent returns a completion and runs none of them (measured 2026-09-02: tool_ids with stream: false executes no tool). Use the chat UI, or bring your own tools with the tools parameter on a model from the list above.

4. Limits

  • Subscription allowance per account (chat + API together): see Usage & metering.
  • Gateway: 64 requests in flight platform-wide, 120 s per request, 2 automatic retries (k8s/litellm/config.yaml, general_settings / litellm_settings). No per-user rate limit is configured today.
  • Uploads, Knowledge collections and notebooks are chat-UI features; the API is text in, text out.