wtf is agentstate?

AgentState is a self hosted proxy server that sits between your AI agent code and the language model it calls, such as OpenAI's GPT-4o. The problem it solves is what happens when an autonomous agent crashes partway through a long task: normally you lose all the progress and have to start over from step one, paying again for every model call along the way. AgentState intercepts every model and tool call, automatically saves the execution state to a local SQLite database as it goes, and lets you resume a crashed run from the exact step it failed on instead of from the beginning. Getting started can be as simple as swapping one import in your Python code for AgentState's own client wrapper, which routes calls through the proxy automatically. There are also ready made wrappers for the LangChain and CrewAI agent frameworks, or you can point any OpenAI compatible client at the proxy's local address without installing anything extra. Once a call has been recorded, replaying that same step on a retry pulls from the cache almost instantly instead of calling the model again, so retries after a crash cost no extra tokens. Beyond crash recovery, AgentState can pause an agent before it performs a sensitive action, such as sending an email or making a payment, and wait for a person to approve or reject it from a dashboard or an API call before letting it continue. It can also automatically reroute requests to a backup model if the primary one is rate limited or returns an error, send Slack or Discord alerts when a run fails or needs approval, and export the full history of an agent's runs as a fine tuning dataset file. A built in dashboard lets you inspect what an agent did step by step and roll it back to any earlier point. Running it locally means cloning the repository, setting up a Python virtual environment, installing dependencies including FastAPI and Playwright, and starting the included server script, which then exposes the proxy on one local port and the dashboard on another. The project is released under the MIT license.