Mole ships a terminal deep-research agent that enforces spend budgets and source-verifies every claim

Mole is a single static binary that turns a research question into a pipeline of decompose, search, extract, verify, and answer steps. The distinctive piece is the budget ledger: each model and search call is reserved against the cap before it runs and reconciled after, with non-negative constraints in the SQLite schema itself, so the ceiling you set is the ceiling you hit. The author reports measured overshoot across their test corpus is 0%. For an engineer running research on someone else's behalf, that turns spend from a tracking problem into a configuration problem, and the budget is a required argument rather than an optional hint.

The other novel piece is the quote gate. During extraction, each candidate claim is checked against the page it was mined from; if the supporting quote is not present verbatim, the claim is discarded before it can reach the answer. Claims that survive can be re-read against their source afterwards, and a claim that turns out unsupported is marked in the report rather than silently dropped. That shifts the workflow from trusting the model to auditing the report, which is a meaningfully different operating mode for research that has to survive review.

For local data, mole accepts a CSV or folder and only returns aggregates such as counts, means, test results, and buckets of at least five records, so individual rows do not leave the machine. A `mole crossings` log records exactly what was emitted, which is the kind of audit trail a security review can actually use. Mole also speaks MCP, so a coding agent can hand it a question and collect the answer, or, in toolkit mode, do its own reasoning while mole supplies the non-model pieces like search and verification. That makes it composable with the agent an engineer already runs rather than a separate console to babysit.

Adopting it costs a search provider key, either Tavily or Brave, and a model provider key, either Anthropic or any OpenAI-compatible endpoint including local servers, both stored in ~/.config/mole/config.json at mode 0600 rather than environment variables. The example config names claude-sonnet-5 and claude-haiku-4-5; switching to a self-hosted model means a token budget rather than a dollar one, since mole only prices search calls in USD. A token-bounded run against a local model costs no money but still counts, so a runaway agent is bounded by tokens. Distribution covers Linux and macOS on amd64 and arm64 via an install script, Homebrew, the AUR.deb, and .rpm, with two name collisions worth knowing: an unrelated `mole` in homebrew/core that is a macOS cleanup tool, and `mole` and `mole-bin` on the AUR held by an SSH tunnelling tool since 2020, both of which the project documents.

Only one feed is carrying this, so corroboration is thin, and the 0% overshoot figure and the verbatim-quote gate are the author's own benchmarks and design notes rather than third-party measurement. The privacy boundary is described, not demonstrated against an adversarial model, and the actual security of the local-data path depends on how the aggregates are computed and which columns the template picks. None of this is unusual for a Show HN post, but it sets the bar for what a second look should measure.