System One Connector MCP
MCP server that gives an agent probabilities and a pick from set options, like ticket urgency, so code can branch on a number
- 361 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The state text is sent to an external API run by TypeSafe, OpenRouter or the owner of the chosen model
- Every call is billed to your key, and items mode multiplies the cost by the number of elements
Install
- Works in
Manual install
curl -fsSL https://raw.githubusercontent.com/itsmostafa/system-one-connector/main/install.sh | shThe macOS and Linux script checks the release archive against its SHA-256 and puts the binary in ~/.local/bin. evaluate update upgrades it.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The state text is sent to an external API run by TypeSafe, OpenRouter or the owner of the chosen model
- Every call is billed to your key, and items mode multiplies the cost by the number of elements
This is third-party code. Review the repository files before installing.
What it does
System One Connector installs a single command, evaluate: a Go binary with no Node or Python runtime. The evaluate setup mcp command finds Claude Code, Claude Desktop, Codex and Hermes and registers an MCP server with one tool, also called evaluate, in each. pi has its own step, evaluate setup pi. The agent sends a state (text or JSON with the facts) and questions of three types. A noul question returns the probability that a condition holds, choice picks one option from a set, and score places the answer on a scale of levels. Every option comes back with a probability, and choice and score also carry a confidence from 0 to 1. The agent compares the number with a threshold and decides whether to act or ask you. The optional min_confidence field enables abstaining: with low confidence the question answers uncertain. The answer comes from TypeSafe's Jev model through the TypeSafe API or through OpenRouter. Profiles (evaluate profile add and evaluate profile use) connect other models: Liquid AI d1, and CLM and Laya hosted on your own machine. In items mode one call asks the same questions about a list of elements, up to 500, and each element goes out as its own request.
Who it is for. People building pipelines on Claude Code, Codex or Hermes where the next step depends on a judgment: how urgent an email is, which team owns a task, what tone a customer reply has.
Good fit when
- Support tickets get routed to teams, and the agent acts on the probability of a choice while handing unsure answers to a person
- Hundreds of reviews, emails or log lines are labelled in one call with the items parameter
- The agent has a threshold: it executes answers with high confidence itself and brings the rest to you
- The model runs on your own server: CLM or Laya connect through the TYPESAFE_BASE_URL variable
Not a fit when
- You need long text, code or reasoning: the tool returns only probabilities and a chosen option
- The data must not leave your network: on cloud routes the state goes to the TypeSafe API, OpenRouter or Liquid AI
- The work machine runs Windows: the install script targets macOS and Linux
- The decision is irreversible and costly: confidence shows how concentrated the probabilities are and does not guarantee the answer is right
Example request
Use evaluate to decide whether this ticket is urgent and which team should own it: "Payouts have been failing for 3 days"MCP
- Transport
- stdio
- Authentication
- API key
Environment variables
TYPESAFE_API_KEYrequired, secret- TypeSafe API key from console.typesafe.ai. Needed unless OPENROUTER_API_KEY is set
OPENROUTER_API_KEYsecret- OpenRouter key for the Decisions route. When both keys are set, TYPESAFE_API_KEY wins
TYPESAFE_BASE_URL- Address of your own server or proxy instead of api.typesafe.ai
TYPESAFE_MODEL- Default model, for example clm-latest
Limitations
A key is required: TYPESAFE_API_KEY from the typesafe.ai console or OPENROUTER_API_KEY. In items mode every element is billed as its own request. The OpenRouter route uses a path under /api/alpha/, and the author warns it may change. The install script works on macOS and Linux. The confidence value describes how concentrated the probabilities are and says nothing about accuracy, so tune the threshold on your own data. Span-01 models accept only noul questions, and running CLM locally needs a GPU.
How to disable. The evaluate setup mcp command registered the server in the clients' settings, so remove the evaluate block from the MCP configuration in Claude Code, Claude Desktop, Codex or Hermes by hand. The binary lives in ~/.local/bin (or the directory from EVALUATE_INSTALL_DIR) and is deleted like any file. The documentation has no uninstall command.
FAQ
How do I connect evaluate to Claude Code?
Install the binary with the install script and run TYPESAFE_API_KEY=your-key evaluate setup mcp. The command finds Claude Code, Claude Desktop, Codex and Hermes and registers the server in each. After you restart the client, ask the agent to call evaluate.
What does evaluate return for an urgency question?
A number from 0 to 1. In the documentation example a ticket about three days of failing payouts scores 0.95. A choice question returns the chosen option, a confidence and a probability for every option.
Why is a bool question rejected?
TypeSafe calls that type noul. The tool refuses bool and any unknown type before sending the request and points to the correct name.
Who pays for the calls?
The owner of the key: requests are billed to your TypeSafe or OpenRouter account, and in items mode every element counts as its own request. The README gives no prices; check them with the provider.
Can it run without a cloud model?
Yes, for CLM and Laya: start a server that implements POST /v1/systemone, put its address in TYPESAFE_BASE_URL and set any non-empty TYPESAFE_API_KEY. CLM needs a GPU.
Discussion
No questions or reviews yet.
Related
MCP serversJev-Style runs small decision models on your machine: a probability per option, six skills, MCP and a Claude Code guard hook
An MCP server giving an agent three Jev tools: typed yes or no, choice and score questions with probabilities, offline by default.
A CLI for any MCP server: list and call tools, resources and prompts, run mock and proxy modes, guard access and manage agent configs
An MCP server that inserts a human feedback step into agent tasks: a web or desktop window for text, images and prompts
An MCP and ACP adapter for the TypeSafe Jev decision model: typed answers and confidence gates for clicks and tool calls.
A local web service and MCP server that gathers several models to answer, peer review each other or debate as advisor personas.
README in short
The page explains the case where an agent needs a quick judgment with a probability and shows install and connection in two commands with a sample question and answer. It goes on to a feature list, a table of models that work behind the System One interface and links to detailed documents: configuration, the tool reference and building from source.