Field notes on the network API of the Tiiny Pocket: every endpoint I've found, what was actually tested, recipes, quirks, and a tool to see what changed after a TiinyOS update.
These are my personal findings, from my own Tiiny, and a work in progress. I'll keep updating them as I learn more about the device and as new features turn up. They are not official documentation and may be wrong or out of date for your TiinyOS version.
Not affiliated with Tiiny. Everything here comes from calling my own device with its own API key, the specs it publishes, Tiiny's public docs and the TiinyOS desktop app. Behaviour can change with any update; each section says which version it was checked against. Corrections welcome (issues/PRs).
| file | what |
|---|---|
| docs/API.md | the reference: services and ports, auth, system/power, models, inference (OpenAI, Anthropic, Ollama styles), model containers, the vault (/kb), connectors, apps, known quirks |
| docs/endpoints/gateway.md, kb.md | every endpoint the device's two published specs list, regenerated by the tool below |
| docs/RECIPES.md | copy-paste: power off, vault file flow, delete captured chats, use the Tiiny from Hermes / Cline / OpenAI SDKs |
| docs/VOICES.md | every voice of every text-to-speech model, and which voice values work |
| docs/HERMES.md | step by step: add the Tiiny as a provider in Hermes Agent |
| docs/MODELS.md | every model in the Tiiny Model Store: type, size, NPU share, context, what it can do |
| docs/PRIVACY.md | what the device keeps about your chats, and what you can delete through the API |
| tools/api_test.py | walk through the API and check each feature really works (chat, embeddings, OCR, images, voice, Jev classification, music, 3D); run one section or all |
| tools/api_compare.py | put api_test reports side by side (e.g. USB vs Wi-Fi, or before/after an update) and mark what differs |
| tools/model_list.py | regenerate docs/MODELS.md from your device's Model Store catalogue |
| tools/cli_compare.py | run the same tests through the Tiiny CLI and the direct API, side by side (plus tiiny bench vs the API's own speed) |
| tools/jev_bench.py + jev_eval_build.py | benchmark the Jev / System One models: accuracy, calibration and speed on a fixed set of public-dataset items (8 tasks, 1,600 items) |
| tools/speed_test.py | speed sweeps: chat reading/writing speed at 1K-128K context, and Jev (/v1/systemone) time per text size and number of questions |
| tools/download_models.py | download models from a script and know when they're done (works around the CLI never returning) |
| tools/tts_voices.py | list each text-to-speech model's voices and check which voice values work through the API |
| tools/usb_download_test.py | check whether the Tiiny can download a model over the USB cable with its Wi-Fi off (turns Wi-Fi back on afterwards) |
| tools/api_snapshot.py | fetch the device's specs, list added/removed/changed endpoints since last time, regenerate the endpoint lists, probe the read-only routes |
See also: AntAttack's documentation map, which covers where Tiiny's own documentation lives and how API-call logging works.
export TIINY_URL=http://192.168.1.50 TIINY_KEY=your-api-key # TiinyOS → Settings → API Key
python tools/api_snapshot.py # first run saves a baseline in snapshots/ (not committed)
python tools/api_snapshot.py --check # later: what changed?Something not working? Test just that part, or everything:
python tools/api_test.py ocr # one section: info chat embeddings ocr image audio classify music 3d
python tools/api_test.py # full run (loads ~10 models in turn; 20-40 minutes)
python tools/api_test.py --list # every check and the model it uses
python tools/api_test.py --via usb # label the run; then --via wifi, and compare:
# api_test.py and speed_test.py stop if a model they use has an update waiting (--allow-outdated to run anyway)
python tools/api_compare.py snapshots/api-test-*/report.json
python tools/cli_compare.py # CLI vs API (needs the Tiiny CLI, signed in)
python tools/usb_download_test.py # over the USB cable: download with Wi-Fi off?
python tools/speed_test.py chat # speed at 1K..128K context (also: speed_test.py jev)It loads the models it needs (and puts back what was loaded before), checks each answer against known inputs in
tools/assets/, and prints PASS/FAIL/SKIP per check; a report and the outputs land in snapshots/.
Don't use the Tiiny for anything else while it runs.
Python 3.10+, standard library only. The key is only sent to the address you give.
- Only things anyone can check with their own device and API key. No root/firmware internals, nobody's addresses, keys or data, and security issues go to Tiiny privately first.
- Mark what was tested versus only seen in a spec or the docs.
License: Apache-2.0.