Mole caps your research agent's spend before it runs, and overshot by zero
Show HN today: Mole, a deep-research agent that lives in your terminal. Written in Go, Apache-2.0, static binaries for Linux and macOS, works against Anthropic, OpenAI, or any OpenAI-compatible endpoint. Forty-nine stars at the time of writing, which is exactly the stage where it is worth looking, because the design decisions are still legible.
Three of them are worth stealing regardless of whether you use the tool. First, the budget is enforced before execution rather than reported after. You set a token or dollar cap and it is hard-capped, with zero measured overshoot across test runs. Every research agent I have run has a story that ends with a bill, and "we'll show you the usage graph afterwards" is not a control. Enforcement at plan time is.
Second, every claim in the output carries a verbatim quote that is checked against its source, and claims that fail the check are discarded rather than softened. Not a citation link. The actual string, verified. That is the difference between a bibliography and a receipt, and it kills the most common failure of research agents, which is a fluent paragraph with a plausible footnote attached to something the source never said.
Third, there is an explicit privacy boundary for local data. It analyzes your files through SQL and only aggregated results ever leave the machine — row contents never go to the model. Same instinct as Google's HEIR announcement today from the opposite end of the sophistication scale: compute over data the model does not get to read.
The pipeline is decompose into sub-questions, run web, academic and local actors, extract and verify claims, detect contradictions between them, then synthesize only from evidence that survived. And it runs as an MCP server, so your coding agent can drive it, or you use toolkit mode where your agent's model does the reasoning and Mole handles the deterministic parts. That split — model judgment separated from deterministic operations — is the same architecture that Spark-to-Paper credited for taking fabrication detection from 14 to 92 percent. Two projects, one insight.
github.com/lajosdeme/mole.
← Back to all articles
Three of them are worth stealing regardless of whether you use the tool. First, the budget is enforced before execution rather than reported after. You set a token or dollar cap and it is hard-capped, with zero measured overshoot across test runs. Every research agent I have run has a story that ends with a bill, and "we'll show you the usage graph afterwards" is not a control. Enforcement at plan time is.
Second, every claim in the output carries a verbatim quote that is checked against its source, and claims that fail the check are discarded rather than softened. Not a citation link. The actual string, verified. That is the difference between a bibliography and a receipt, and it kills the most common failure of research agents, which is a fluent paragraph with a plausible footnote attached to something the source never said.
Third, there is an explicit privacy boundary for local data. It analyzes your files through SQL and only aggregated results ever leave the machine — row contents never go to the model. Same instinct as Google's HEIR announcement today from the opposite end of the sophistication scale: compute over data the model does not get to read.
The pipeline is decompose into sub-questions, run web, academic and local actors, extract and verify claims, detect contradictions between them, then synthesize only from evidence that survived. And it runs as an MCP server, so your coding agent can drive it, or you use toolkit mode where your agent's model does the reasoning and Mole handles the deterministic parts. That split — model judgment separated from deterministic operations — is the same architecture that Spark-to-Paper credited for taking fabrication detection from 14 to 92 percent. Two projects, one insight.
github.com/lajosdeme/mole.
Comments