Runcrate assistant for Tabularis

A declarative ai-assistant package that lets Tabularis generate SQL with the open models on Runcrate, an AI cloud. Runcrate serves open models through one OpenAI-compatible API at https://api.runcrate.ai/v1, billed per token from prepaid credits. This package is maintained by the Runcrate team (Aeonmind LLC).

The package contains JSON and an icon; Tabularis owns HTTP, credentials, proxy routing and prompt construction. No plugin code executes inside the application.

Requires Tabularis 0.32.1-4 or newer (nightly channel until a compatible stable release is available). Older runtimes are blocked by the package's min_runtime_version. The AI assistant runtime was introduced by Tabularis PR #1337.

Install and configure

1. Create an API key (how) and add prepaid credits. Requests are billed per token.
2. In Tabularis open Settings → AI → Available providers → Runcrate → Install and use.
3. Enter the key, or set RUNCRATE_API_KEY in the environment running Tabularis.
4. Select a model and use AI → Generate SQL in the editor.

How it works

• Models. Discovery reads the public GET /v1/models endpoint without credentials and keeps the models whose output_modalities include text. The list is cached for 24 hours. These preferred models are listed first: deepseek-ai/DeepSeek-V4.1-Flash, zai-org/GLM-5.3-Flash, zai-org/GLM-5.3, Qwen/Qwen3.8-Flash, openai/gpt-oss-120b and moonshotai/Kimi-K2.6. Model IDs are the ones the API returns, including the organization prefix.
• Generation. One POST /v1/chat/completions request per prompt, with the Tabularis system prompt as the system message, the user prompt as the user message and temperature set to 0. The host sends the key as Authorization: Bearer <key>, which the Runcrate API accepts. The answer is read from /choices/0/message/content, so a reasoning model's separate reasoning_content is never mixed into the SQL.
• Scope. Nonstreaming text generation only. The package does not provide OAuth, streaming, tools or multimodal input.

Prices and the current model list are published at api.runcrate.ai/v1/models and in the docs; this package does not copy them.

Package

python3 build.py --output dist
python3 -m pip install jsonschema
python3 -m jsonschema -i .tabularium manifest.schema.json

The ZIP root contains .tabularium, icon.svg, README.md and LICENSE. Releases publish the ZIP plus the standalone .tabularium; the registry resolves the manifest from the release assets and the ZIP is the universal asset. The tag must match the manifest version. Never put API keys in this repository or a package.

manifest.schema.json is a copy of the schema in TabularisDB/tabularis.

License

Apache-2.0. Runcrate is operated by Aeonmind LLC.