What it is
System One Connector, by Mostafa (itsmostafa), is a single Go binary called evaluate that registers one MCP tool, also named evaluate. The agent sends a piece of state and a set of typed questions (yes/no, pick one option, or rate on a scale) and gets back answers with probabilities it can compare against a threshold.
Its point is provider independence. Jev from TypeSafe is the default, through a TypeSafe or OpenRouter key, but the same tool can target any server that speaks the System One API: a self-hosted Laya or CLM, Liquid AI's hosted d1, Jared Palmer's Kev 4B or Respan's Span-01, saved as profiles and switched with evaluate profile use. It can judge up to 500 records in one call, and a per-question min_confidence lets the agent abstain on unsure answers.
evaluate setup mcp finds and configures Claude Code, Claude Desktop, Codex and Hermes, and evaluate setup pi covers pi. The repository has about 350 GitHub stars and an MIT licence; version 0.4.12 was released on 9 October 2026.
Install
curl -fsSL https://raw.githubusercontent.com/itsmostafa/system-one-connector/main/install.sh | sh && evaluate setup mcpOur review
System One Connector does one thing: it passes typed questions to a decision model and returns the probabilities intact. That restraint is its strength. One tool costs little context, the static Go binary needs no Node or Python runtime, and the same call works whether the answer comes from TypeSafe's hosted Jev or a Laya instance on your own GPU.
The details show care. The connector rejects malformed questions locally, including two inputs its tool reference says the Jev API mishandles (an unknown type gets only 'Invalid request', and stray noul criteria keys are silently dropped). It enforces limits of 255 options per choice and 10 levels per score before sending, retries up to three times on 429 and 529 errors, and documents how to phrase state as evidence rather than conclusions, which matters because a leading field skews the confidence that comes back.
Items mode handles up to 500 records per call with 8 requests in flight, and a failed record does not sink the batch. The catch is on the other side of the call: 500 items return roughly 100 to 300 KB, tens of thousands of tokens that land in the calling model's context. Batch size is a context decision as much as a throughput one.
There are no ready-made checks here: no injection screen, claim verifier or diff gate, which Jev MCP supplies for coding agents. The 'any model' promise is also only as good as the models behind it; Span-01 accepts only yes/no questions, self-hosting Laya or CLM means running a separate server, and the connector verifies none of the providers' accuracy claims.
Where it earns its place
A support-ops engineer running a Claude Code agent over a ticket backlog
Route and prioritise 500 tickets in one pass
Pass the tickets as items with a noul question for urgency and a choice for department; min_confidence turns unsure answers into __uncertain__ for a person to check. At Jev's $42 per billion input tokens the decisions cost cents; the bigger cost is the 100 to 300 KB result in the agent's context.
A health-tech team whose data rules forbid a hosted judging model
Keep yes/no and scoring calls on their own GPU
Point TYPESAFE_BASE_URL at a self-hosted Laya or CLM server and save it as a profile; the agent's tool calls do not change. They must run and size that model server themselves, since the connector ships none.
An ML engineer comparing decision models for a classification step
Run the same questions against Jev, d1 and Kev 4B
Save each provider as a profile and switch with evaluate profile use; probabilities come back unaltered apart from ordering, so outputs compare directly. Span-01 handles only noul questions, so not every model can run every test.
Before you commit
- Setup
- Run the install script with curl (macOS and Linux), then TYPESAFE_API_KEY=your-key evaluate setup mcp, which registers the binary with Claude Code, Claude Desktop, Codex and Hermes where it finds them. evaluate update upgrades the binary in place.
- Integrations
- One tool, evaluate, with three question types (noul, choice, score), an items batch mode and per-question min_confidence. Providers: TypeSafe or OpenRouter for Jev, Liquid AI's hosted d1, Kev 4B and Span-01 through OpenRouter, and self-hosted CLM or Laya through a custom base URL.
- Data and privacy
- The state the agent puts in a call goes to the chosen provider: a third party for hosted Jev or d1, your own network for self-hosted models. The connector fetches nothing itself and sends only what the agent passes; no security advisories are published.
- Cost at scale
- The connector is free. TypeSafe lists Jev at $42 per billion input tokens; the tool reference's sample call used 349 input tokens, so a million similar decisions is about 349M tokens, roughly $15 before output tokens, whose price TypeSafe's homepage does not show.
- Limits
- At most 500 items per call, 8 concurrent requests, 16 MiB of responses per batch and a 60-second request timeout. A choice is capped at 255 options and a score at 10 levels; the quick-install script covers macOS and Linux only.
What's good
- One tool, so very low context overhead
- Swaps between Jev, Laya, CLM, d1, Kev 4B and Span-01 with profiles
- Batch mode for up to 500 records per call
- Single static binary; one command configures several clients
What's not
- No ready-made guards, verifiers or review tools
- Quick install pipes a script from GitHub to the shell
- Self-hosted models need their own server set up separately
- Large batches can flood the calling model's context
Best for
Developers who want raw, typed decisions in their agent and may switch between hosted Jev and a self-hosted open model.
Alternatives
System One Connector: common questions
- Is System One Connector worth it?
- A thin, careful connector that returns any System One model's probabilities untouched and leaves the hard part, writing good questions, to the agent. We rate it 4.1 out of 5.
- How much does System One Connector cost?
- The connector is free (MIT). Model calls are billed by the provider you choose; TypeSafe lists Jev at $42 per billion input tokens. Self-hosted Laya or CLM cost only your hardware.
- What are the downsides of System One Connector?
- No ready-made guards, verifiers or review tools; quick install pipes a script from GitHub to the shell; self-hosted models need their own server set up separately; large batches can flood the calling model's context.
- Who is System One Connector best for?
- Developers who want raw, typed decisions in their agent and may switch between hosted Jev and a self-hosted open model.
- How much does System One Connector cost for a team or in production?
- The connector is free. TypeSafe lists Jev at $42 per billion input tokens; the tool reference's sample call used 349 input tokens, so a million similar decisions is about 349M tokens, roughly $15 before output tokens, whose price TypeSafe's homepage does not show.
- Is System One Connector safe for company data?
- The state the agent puts in a call goes to the chosen provider: a third party for hosted Jev or d1, your own network for self-hosted models. The connector fetches nothing itself and sends only what the agent passes; no security advisories are published.
- What are the best System One Connector alternatives?
- The closest alternatives we have reviewed are Jev MCP, Laya MCP Server and Jev on Viorant.
Answers checked Oct 23, 2026.
Make System One Connector? Show this rating on your site
