Local AI assistants sharing tables, notes and tasks, with scheduled runs and allow, ask or deny rules. Includes the Tholos-2B tool-use model.
See the codeTholos runs small AI assistants that work on a schedule and share tables, notes and a task board. It brings recurring research and monitoring into one local workspace, with rules that allow, ask about or deny each action.
See Tholos or install it below.
Requires Python 3.11 or newer and a model server. For the default Tholos-2B model, install Ollama first, then run:
ollama pull hf.co/mertkayacs/Tholos-2B-GGUF:Q4_K_M && uv tool install git+https://github.com/mertkayacs/tholos && tholos
Open http://127.0.0.1:7070. In Settings, press Detect and add the model. In Starter teams, load Research desk: Scout gathers articles, Analyst scores them and Writer produces a weekly brief. The Price watch team checks product pages and adds a task when a price reaches your target.
The app is one Python process with a SQLite database. The default model's Q4_K_M file is about 1.6 GB and runs on a CPU.
Each assistant has its own role, model, schedule and tools. It can edit tables and notes, search the workspace, hand tasks to teammates and ask you questions. The timeline records every tool call; workspace edits keep their author and version.
Rules allow an action, hold it for approval or deny it. Fetching a web page asks you first by default. Approval applies to the exact arguments shown to you. Conflicting writes require the assistant to read the latest version before updating it.
Tholos validates each model step against the assistant's tool schema. It supports llama.cpp, Ollama and OpenAI-compatible chat completion servers. Tholos-2B documents model-server settings and runtime differences.
Tholos-2B is a 2B tool-use model fine-tuned from MiniCPM5-2B on executed Tholos tasks. On Tholos-Bench, it passed 137 of 160 scenarios, compared with 112 for MiniCPM5-2B. Both ran on a Kaggle T4 with Q4_K_M weights, llama.cpp and JSON schema decoding.
| Model | Passed, of 160 |
|---|---|
| Qwen3.5-2B | 97 |
| MiniCPM5-2B | 112 |
| Tholos-2B | 137 |
| Llama 3.2 3B Instruct | 37 |
| Granite 4.2-3B | 129 |
| Qwen3.5-4B | 141 |
Only Tholos-2B was fine-tuned for the step format. These scores measure workspace tasks in Tholos. Scores differ by runtime: Tholos-2B passed 134 with CPU llama.cpp and 136 with Ollama. Compare models within the same runtime and settings.
The model card gives the benchmark conditions and limits. GGUF weights, training data and training code are available separately.
Tholos listens on loopback by default. Keep it there for local use. It checks request hosts, same-origin requests and form CSRF tokens. Retrieved pages are labelled as untrusted input; web fetches are restricted to public HTTP and HTTPS addresses on ports 80 and 443. Only you can change approval rules.
The database stores API keys and the Telegram token, with file access restricted to your user; the UI masks them. Review approvals and model-server settings before enabling a recurring team.
An Eschatia Labs project by Mert Kaya.
Local AI assistants sharing tables, notes and tasks, with scheduled runs and allow, ask or deny rules. Includes the Tholos-2B tool-use model.
See the codeTholos runs small AI assistants that work on a schedule and share tables, notes and a task board. It brings recurring research and monitoring into one local workspace, with rules that allow, ask about or deny each action.
See Tholos or install it below.
Requires Python 3.11 or newer and a model server. For the default Tholos-2B model, install Ollama first, then run:
ollama pull hf.co/mertkayacs/Tholos-2B-GGUF:Q4_K_M && uv tool install git+https://github.com/mertkayacs/tholos && tholos
Open http://127.0.0.1:7070. In Settings, press Detect and add the model. In Starter teams, load Research desk: Scout gathers articles, Analyst scores them and Writer produces a weekly brief. The Price watch team checks product pages and adds a task when a price reaches your target.
The app is one Python process with a SQLite database. The default model's Q4_K_M file is about 1.6 GB and runs on a CPU.
Each assistant has its own role, model, schedule and tools. It can edit tables and notes, search the workspace, hand tasks to teammates and ask you questions. The timeline records every tool call; workspace edits keep their author and version.
Rules allow an action, hold it for approval or deny it. Fetching a web page asks you first by default. Approval applies to the exact arguments shown to you. Conflicting writes require the assistant to read the latest version before updating it.
Tholos validates each model step against the assistant's tool schema. It supports llama.cpp, Ollama and OpenAI-compatible chat completion servers. Tholos-2B documents model-server settings and runtime differences.
Tholos-2B is a 2B tool-use model fine-tuned from MiniCPM5-2B on executed Tholos tasks. On Tholos-Bench, it passed 137 of 160 scenarios, compared with 112 for MiniCPM5-2B. Both ran on a Kaggle T4 with Q4_K_M weights, llama.cpp and JSON schema decoding.
| Model | Passed, of 160 |
|---|---|
| Qwen3.5-2B | 97 |
| MiniCPM5-2B | 112 |
| Tholos-2B | 137 |
| Llama 3.2 3B Instruct | 37 |
| Granite 4.2-3B | 129 |
| Qwen3.5-4B | 141 |
Only Tholos-2B was fine-tuned for the step format. These scores measure workspace tasks in Tholos. Scores differ by runtime: Tholos-2B passed 134 with CPU llama.cpp and 136 with Ollama. Compare models within the same runtime and settings.
The model card gives the benchmark conditions and limits. GGUF weights, training data and training code are available separately.
Tholos listens on loopback by default. Keep it there for local use. It checks request hosts, same-origin requests and form CSRF tokens. Retrieved pages are labelled as untrusted input; web fetches are restricted to public HTTP and HTTPS addresses on ports 80 and 443. Only you can change approval rules.
The database stores API keys and the Telegram token, with file access restricted to your user; the UI masks them. Review approvals and model-server settings before enabling a recurring team.
An Eschatia Labs project by Mert Kaya.