Local-first, deterministic, never calls an LLM. Your coding tool does the thinking — Visp handles scope, evidence, and memory.
$ npm install -g visp-kit
We don't add another AI to your stack — we make the AI you already use provable.
You point an AI coding tool at a task. It edits ten files when you asked for two, writes a test that passes whether or not the fix works, and hands you a diff you now have to fully re-read to trust.
src/auth/session.ts
in scope
src/auth/token.ts
in scope
src/auth/guards.ts
in scope
src/billing/plan.ts
out of scope
Declare before editing. Block outside the boundary. Record evidence you can check without reading every line.
Deterministic and local-first — the boundary is enforced by the tool, not negotiated with a model.
Stops AI-written code reaching review without proof.
Makes the tool declare what it will change before it edits, blocks anything outside that scope, and records evidence you can check without reading every line. Never calls an LLM.
$ npm install -g visp-kit
Runs the Visp workflow inside the AI coding tool you already use.
Sits inside Codex, Claude Code, Copilot, or OpenCode. Prepares each task, hands the agent only the context it needs, records what happened. Never calls an LLM itself.
$ npm install -g visp-hyper-agent
Curated, auditable memory for coding agents.
Mines your existing git history so memory is useful in the first minute, not after weeks of note-taking. Deliberately careful about how little it injects.
Most memory tools optimise for injecting more. SWE-ContextBench (1,100 tasks, 51 repositories) found well-chosen prior context lifts issue resolution — but context the agent retrieved for itself, unfiltered, scored 12.1 points below the curated version and cost more tokens than using no memory at all.
$ npm test Test Files 199 passed (199) Tests 1209 passed (1209) Duration 33.38s
Vispnote, our desktop notes app, is built on these tools — vispnote.com ↗