# Ekbasis [Health: Active]

**Category:** 💻 Developer Tools  
**Repository:** https://github.com/OpenInterpretability/ekbasis  
**GitHub Stars:** 19  
**Views:** 0  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/ekbasis

## Description
Foresee what git commands and multi-step changes will do, as calibrated probabilities

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "ekbasis": {
    "command": "uvx",
    "args": ["ekbasis"],
    "env": {
      "EKBASIS_URL": "",
      "EKBASIS_API_KEY": "",
      "EKBASIS_SHELL_GUARD": ""
    }
  }
}
```

**Requires environment variables:** `EKBASIS_URL`, `EKBASIS_API_KEY`, `EKBASIS_SHELL_GUARD` — the values above are empty placeholders; fill in real credentials before running (see the repository for what each one is for).

## Documentation

## What Ekbasis MCP server does

Ekbasis MCP server provides an action-conditioned consequence model for AI agents. Given a description of the current state and a proposed action, it estimates what the next state will look like. Questions can cover whether a command will fail, whether work will be lost, or what a value will become after the action. Results are calibrated probabilities produced in one forward pass rather than generated text.

The model is intended to sit between an agent’s plan and execution. It can provide a quick first assessment, allow an agent to escalate uncertain cases to a more expensive reasoning step, or be chained one action at a time to track a longer sequence. Training includes synthetic environments and real executions of git commands in temporary repositories.

## How it works

The client sends state, action, and question information to an Ekbasis model endpoint. The endpoint may be the hosted API at `https://openinterp.org/api/v1` or a service operated locally. A self-hosted deployment uses the published open weights with vLLM and requires a GPU with approximately 80 GB of memory.

For git workflows, the repository state is represented in the format used during training, and commands are checked before they run. The Claude Code integration adds a fail-closed hook: if a check cannot complete, such as because of a hook error or a response timeout, it asks for confirmation rather than allowing the command through automatically.

## Setup and configuration

Install the Python client with `pip install ekbasis`. Python 3.9 or newer is required for the client. To add the MCP server, install `ekbasis[mcp]` with Python 3.10 or newer, or run it through `uvx` with the `mcp` dependency. The standard MCP command is `uvx --python ">=3.10" --with "mcp>=1.2" ekbasis mcp`.

Set `EKBASIS_API_KEY` to use the hosted service; a key beginning with `ekb_` can be obtained from the Ekbasis console. Alternatively, set `EKBASIS_URL` to the hosted endpoint or to your own compatible service. Run `ekbasis health` to check connectivity. Claude Code users can install the repository plugin, which includes the hook, skill, MCP server, and command-line client; it requires `python3`, `sh`, and `uv`.

## Tools and capabilities

- Ask consequence questions through the client’s `ask` workflow.
- Check individual git commands with `ekbasis git-check`.
- Evaluate a batch of planned changes with `ekbasis plan-check` and a rules file describing the system.
- Test plan-check rules against known cases with `--cases`.
- Use the MCP server from MCP-compatible clients.
- Run a Claude Code git guard that checks commands before execution.

## Limitations and notes

Ekbasis MCP server does not act as a general reasoning model or produce open-ended text answers. Its usefulness depends on how accurately the current state, action, question, and any plan-check rules describe the system. The git guard is specialized for git; the shell guard outside git is described as a prototype and may ask too often.

Hosted use requires an API key or endpoint configuration. Self-hosting involves downloading the model, running vLLM, and starting the supplied serving process. The model produces predictions, not guarantees, so uncertain or high-risk results should be handled by the surrounding agent or an explicit confirmation step.

_Full upstream README: https://allmcps.com/mcp/ekbasis/readme_