activegraph-monid

activegraph-monid is a Python research agent that takes a plain-language question or topic, breaks it into strategies and tasks, queries multiple data sources to gather information, and produces a written memo summarizing what it found, with citations. The whole process runs automatically once you give it a starting goal.

The agent works by chaining together eleven steps. First it uses a language model to decompose the goal into strategies and then into specific queries. For each query it consults a service called Monid to discover which data endpoints might have relevant information, then picks one, inspects its input requirements, builds a valid request, and runs it. The results are passed back to the language model to extract specific claims, evaluate whether enough information has been gathered, and finally write a memo that cites where each claim came from. The confidence level for each claim is shown in the output.

Everything that happens during a run is recorded as a typed event in a local SQLite database called a trace. This makes the run auditable: you can inspect the full chain of decisions, see which sources were queried and why, and understand exactly how the final memo was produced. The trace also supports forking, where you can swap out one data source and replay only the steps that depend on it, rather than rerunning the whole thing from scratch. This lets you compare memos from different source combinations cheaply because LLM and tool call results are cached.

The agent accepts a cost budget in US dollars and a limit on how many data endpoints it may use, so you can cap spending before starting a run. The README shows an example run about a company's funding history that cost roughly a dollar in total across both the Monid API and the language model API calls.

Setup requires Python 3.14 or later and the uv package manager. You supply an Anthropic API key and a Monid API key in a config file, then run the agent from the command line.