OY1 / CHATGPT PLUGIN
Your ChatGPT.
Working smarter.
OY1 is a ChatGPT plugin built to help it remember your task, check its work and learn from corrections. Connect once, then use OY1 in your chats.
01 / THE HARNESS
Structure around the model.
OY1 adds memory and a checking workflow around ChatGPT’s model. The connection uses the Model Context Protocol (MCP), so other compatible AI apps can use it too. Your AI app handles reasoning and external actions; OY1 keeps the task history and checks the results it reports.
The hosted OY1 MCP service supplies nine tools for starting a task, proposing a step, recording an observation, saving and recalling memory, planning from recorded transitions, finishing or cancelling, and checking the connection.
A model generates a response; a harness manages the surrounding task; MCP connects the client to tools. OY1 supplies the harness through MCP. Your client still chooses the model and performs external actions.
02 / THE WORKFLOW
Predict. Act. Check.
- Define what success means.
The client begins with your goal and concrete success criteria. A saved task ID ties the work together.
- Record the next step.
Before acting, the client supplies a hypothesis, an action and testable predictions. Only one action can be pending at a time.
- Observe what happened.
The client uses its own tools, then reports the result. OY1 compares observations with predictions and requires a revision after a mismatch.
- Finish with recorded checks.
Completion requires a passing check for every agreed criterion and no pending action. Uncertain outcomes must be resolved before continuing.
03 / MEMORY
Resume with context.
Task memory distinguishes hypotheses, supported findings and refuted ideas. Supported memories reference recorded observations. Recall returns the task’s current state, version and history, so the client can continue after losing context without replaying previous actions.
The planner searches unambiguous transitions already recorded in the task. It does not simulate unknown outcomes. Task state is saved on the OY Labs service and survives service restarts.
Completed and cancelled tasks and their history are automatically deleted after 90 days. Unfinished tasks expire after 180 days without changes. You can download or delete any saved task from your account.
04 / QUESTIONS
Before you connect.
Does OY1 include a language model?
The hosted MCP service uses the model access you already have in your AI client. OY Labs does not make model inference calls for this service. Model access and any related charges remain with your client or model provider.
Do I need to host the harness myself?
No. OY Labs hosts the MCP service and its task controller, memory and planner. Connect a compatible client to https://oylabs.ai/mcp; a local runner is not required for this hosted service.
How do I connect?
You need: a Mac or Windows computer, ChatGPT Plus or higher, and the ChatGPT desktop app. Enable Computer Use in ChatGPT Work, then send the setup prompt.
What does a passing check prove?
It shows that the client reported a result satisfying the task’s recorded criteria. OY1 does not independently authenticate external facts, force a client to use its tools, or guarantee a correct answer. Review consequential results yourself.
Is this the same as the ARC-AGI-3 benchmark runner?
This hosted MCP adapter uses the original harness memory and planner with a separate task controller. It does not execute the frozen ARC benchmark runner or visual perception pipeline. Benchmark results should not be treated as a measurement of this adapter on general tasks.
Can I see the source code?
Yes. Our ARC-AGI-3 API harness is open source on GitHub, including code, results and reproduction instructions. That repository covers the benchmark harness; the ChatGPT plugin connects to the service hosted by OY Labs.
Put OY1 to work.
Start with the connection guide, explore the tool reference, and approve the connection with your OY Labs account.
Connect your client →