VinvAI

Why Vinv

84% of developers now use or plan to use AI coding tools — but more of them distrust the output (46%) than trust it (33%), and distrust nearly doubled in a year (Stack Overflow 2025, 49k developers).

The reason is familiar: the agent edits the wrong handler, invents return shapes, then grades its own homework while the server won't even start. Or it loops — test fails, agent edits the same function, test fails again — burning your context window on "let me verify."

Both failures share one root cause: the agent has never watched your code run. It argues from static text.

The industry automated *writing* code and left *proving* it entirely manual. Vinv automates the proving — and only then the finding and the fixing.

Install

Vinv works three ways — an editor extension, a CLI, or an MCP server for any agent. All three share the same local engines.

Editor extension

One click — **install from vinv.ai**, which opens the

extension directly in your editor (VS Code, Cursor, Windsurf, VSCodium, Trae and

Insiders). The listing lives on

Open VSX and the

VS Code Marketplace.

Or from your editor's CLI:

First run builds the engines — about 3 minutes, mostly compiling the Rust index; it also fetches a one-time ~100 MB local embedding model (uv and Rust required). First trace lands about a minute after that; everything after is seconds.

CLI (Python engines)

pip install vinv          # every engine as a console script
# or run one with zero install:
uvx --from vinv exerciser campaign  --budget 20

MCP server (any agent)

Give Claude Code, Cursor, or any MCP client Vinv's tools — one global config that finds your open workspace automatically via MCP roots.

Claude Code, Codex, Gemini CLI — installed once for every folder you open, not

just the current one:

claude mcp add --scope user vinv -- npx -y vinv-mcp

codex mcp add vinv -- npx -y vinv-mcp

gemini mcp add --scope user vinv npx -- -y vinv-mcp

Claude Code defaults to the current directory and Gemini CLI to the current project, so both take --scope user; Codex always writes to ~/.codex/config.toml and has no scope flag.

Other clients: add { "command": "npx", "args": ["-y", "vinv-mcp"] } under mcpServers.vinv. See — 16 tools: semantic search, dead code, fault localization, runtime values/slices/coverage, and the verify/optimize loop.

Build from source — for contributors

git clone https://github.com/VinvAI/VinvAI ~/.vinv/engines && cd ~/.vinv/engines && ./install.sh

Windows (PowerShell):

git clone https://github.com/VinvAI/VinvAI $HOME\.vinv\engines; cd $HOME\.vinv\engines; .\install.ps1

The loop, in one picture

Run → Test → Find → Prove. Point Vinv at a Python repo; it does the rest — no code changes, no API keys.

flowchart LR
  T[Trace] --> I[Index] --> S[Serve MCP] --> V[Verify] --> L[Learn] --> T

• Run — brings every service up under tracing with zero edits, capturing timings, arguments, return values and call trees from the real run.
• Test — drives real requests through every endpoint (valid, boundary, negative, authenticated) and banks each response as a regression case.
• Find — surfaces what actually broke or slowed down — server errors, crashes, latency hotspots, dead code — each tied to the exact source line.
• Prove — hands the evidence to the agent you already use, then verifies its fix against acceptance tests written *before* the fix that it never sees. A "faster" change that alters any output is auto-reverted.

Your agent is the only LLM — no new bill, no model picker, no provider keys. See the full walkthrough on .

See it in action

Not a lab benchmark — real findings, filed on real projects (scikit-learn, watermarks-remover, semantica, FastAPI, Typer, smolagents), every one with an upstream thread you can open. All of it driven by Cursor running Composer 2.5 with Vinv installed — not a frontier model. The evidence did the work, not the model. The full set with screenshots: **vinv.ai/#catches**.

Dead code — proven by what never ran

Static tools only prove *"nothing references this."* Vinv proves *"no capture ever executed this,"* carries each untraced island with the live callers that still point at it, and lets your agent return the verdict — integrate, delete, or keep.

semantica#1176 is the discipline in one thread: a maintainer flagged two symbols as importable downstream, Vinv restored them with deprecation warnings, and the same maintainer merged.

Optimization — a speedup that has to prove itself

Every call is timed and charged to the symbol that spent it. A candidate fix ships only if a paired-bootstrap 95% CI clears zero *and* the behavior suite replays byte-identical — faster-but-different is auto-reverted.

Bug report — bugs that only exist while something runs

Scanners read source and guess. Vinv drives the service and watches what comes back:

• drive — every discovered endpoint, exercised with inputs nobody wrote.
• judge — a 500 is not a finding on its own; the oracle names the status that *should* have come back.
• hand over — ships as an evidence pack: repro command, caller chain, real argument values.

On fastapi/full-stack-fastapi-template (~44k★), the authenticated sweep filed **discussion #2454** — four endpoints answering 500 to input that should be 4xx, one repro each — then another contributor reproduced all four against `master`, file-and-line: *"checked against master — they're all real."*

What Vinv gives your agent

One loop, grouped by what you came to fix. Each capability links to a live walkthrough on the site.

Understand the codebase → vinv.ai/#what-it-does

• Semantic code search — ask by meaning, get ranked symbols with `def` bodies and line numbers, embedded by a local model.
• Code Graph — a persistent map of every symbol and call edge, updated incrementally on save, with a live runtime overlay.
• Runtime tracing — zero-edit tracing: timing, memory, args, returns, errors, per call, joined to source.
• Ask Vinv — plain-English questions about your running system, every answer citing the exact trace spans and source lines, with a deterministic critic that blocks any claim the evidence can't back.

Find what actually broke → vinv.ai/#catches

• Rank suspects — on any failure, symbols ranked by fault-localization score over real pass/fail requests, with the real error messages attached.
• Behavior exerciser — Vinv doesn't wait for traffic: it drives every discovered endpoint itself, picks strategies with a Thompson-sampling bandit, and banks every response as a permanent regression case.

Clean up dead code → vinv.ai/#what-it-does

• Dead-code sections — untraced *islands* split into "no references" vs "reached from live code but never taken," each with the live callers that still point at it and a keep-or-cut verdict with reasoning.

Recover latency → vinv.ai/#what-it-does

• Recoverable time — latency hotspots ranked by the milliseconds you'd actually get back, each dispatched as a predicted-then-proven optimization instead of a guess.

Trust the fix → vinv.ai/#under-the-hood

• Verified fixes — replayed start, live port, acceptance tests the agent never sees. One click reverts everything an episode touched.
• Journey — one walkthrough of everything verified: every service, then every endpoint's call tree, latency flamegraph, and exact inputs → outputs, with a form to add your own test inputs that the engine replays forever.
• Auto-Pilot — one click drives discover → set up → trace → exercise → fix → verify, until green or budget; when new trace errors land, the fix episode is already dispatched by the time you see the red ring in the graph.
• Findings — issue clusters, optimization episodes with paired-bootstrap confidence intervals, regression diffs by kind, and a machine-readable `findings.json` your agent can consume directly.

> Honest scope: Python first — services and APIs. Other languages get the index, graph and grounded Q&A, but no runtime evidence yet. TypeScript and Go are next.

Agent alone vs Agent + Vinv

Why agents can't reward-hack under Vinv

Vinv ties every runtime trace to the exact code that produced it and hands your agent a context graph built from that join — so the agent argues from evidence, not vibes. And when it claims victory, Vinv doesn't take its word:

• **Acceptance tests are authored *before* the fix**, stored outside the workspace under an opaque token, and must fail deterministically twice on the broken code — a test that passes pre-fix is thrown away.
• A "faster" fix that changes any observable output is auto-reverted — the behavior suite must replay byte-identical and the speedup's paired-bootstrap 95% CI must exclude zero.
• Deliberate 4xx rejections aren't "errors" to fix — the defect classifier knows a service saying *no* correctly from a service breaking, so the agent is never handed a fake goal it can only game.
• When two attempts stop making progress, a Nash-bargaining stall judge continues only if both an explorer stance *and* an auditor stance strictly prefer it to asking you.

Under the hood: the oracle roster, the budget dispatcher, and the sandbox

The Test stage isn't one tester — it's a set of oracles, each hunting a different class of defect, all writing into the same findings and fix-dispatch path.

The dispatcher is a bandit. `exerciser campaign` allocates one budget across every *armed* oracle by Thompson sampling over `(target × technique × oracle)`. Cost is *measured* (wall-clock normalized to probe-equivalents plus subprocesses spawned), and credit is paid once per defect signature so a deterministic oracle can't re-earn credit for the same bug. Posteriors persist in `campaign.json` — which technique pays *on your repo* is learned.

Unverified code runs behind a containment ladder: a kernel-enforced OS sandbox (`sandbox-exec` / `bwrap` / `unshare`) where the host offers one, otherwise a process shim — always with a disposable repo copy, redirected `HOME`/`TMPDIR`, blocked network and subprocess spawning. The tier is decided by a *probe* that verifies a write outside the root really failed, never by a binary being on `PATH`. Postgres, Redis and S3 are substituted *inside* the jail so code that needs them runs instead of failing to connect.

Proven on itself

Vinv's release gate is Vinv — these numbers come from running the loop on this repository:

The full learning walk — reward, propensity, gating math, with file:line for every claim — is . The test ontology is .

Works with your agent

One command (Register Vinv MCP in Agent Tools) writes the servers into every agent it detects:

Where the config lands, per client — and how to verify

Registration is idempotent and never commits secrets. The servers (`vinv-index`, `vinv-runtime`, `vinv-exercise`) launch over stdio via the editor's own runtime.

• Claude Code — `~/.claude.json`, project-local scope. Verify: `claude mcp list`.
• Cursor — `/.cursor/mcp.json`. Verify: Settings → MCP shows them green.
• Codex CLI — `~/.codex/config.toml` under `[mcp_servers.vinv-index]` / `[mcp_servers.vinv-runtime]`.
• Copilot Chat — native VS Code MCP provider (auto), `.vscode/mcp.json` on older builds.
• Windsurf Cascade — `~/.codeium/windsurf/mcp_config.json`.
• Gemini CLI — dispatch works out of the box; for MCP tools, add the same stdio servers to `~/.gemini/settings.json`.

Your agent is also Vinv's only LLM — every analysis step routes through the coding-agent CLI you already pay for.

MCP tools reference — few tool names on purpose; the session tool multiplexes

`vinv-index` — your code and the session:

`vinv-runtime` — the captured runs (read-only, provenance-stamped):

`vinv-exercise` — your agent exercises your service and reports the run back; Vinv grades what came back.

The engines

`pip install vinv` ships every engine as a console script — or use any one standalone:

exerciser CLI — the oracle swarm, runnable standalone

Requires `identification consolidate` first for `apis.json`, and — for real coverage — a service running under `tracelens`.

FAQ

Do I need my own API keys?

No. Vinv runs everything locally. The semantic index and code embedder run on your machine without any provider keys. Your agent CLI (Claude Code, Cursor, …) handles its own LLM communication using the auth you already set up.

Is there any telemetry or data collection?

No telemetry, no analytics, no usage pings, no crash reports. Vinv stores per-repo state in .vinv/ and per-machine state in ~/.vinv/; sensitive data in traces is redacted and never sent anywhere. The extension makes exactly one outbound request of its own: a GET of a static file at notices.vinv.ai on activation, so a broken release can tell you. No query string, no identifiers, nothing uploaded; at most once an hour. Turn it off with vinv.notices.enabled.

Does Vinv modify my code?

No. Vinv uses a zero-edit tracer, instrumenting your Python backend at runtime without SDK integrations, decorators, or source changes.

How does it know if a fix worked?

Acceptance tests are authored before the fix, stored outside your workspace under an opaque token, and required to fail deterministically twice on the broken code — a test that passes pre-fix is discarded. The fix must then pass them, with a replayed start, a live port, and every other observable behavior byte-identical. A deterministic anti-cheat audit over the diff blocks test edits, swallowed exceptions and shadow modules outright.

Which languages and agents are supported?

Runtime evidence — tracing, the oracle swarm, verified fixes — is Python today, for services and APIs; other stacks get the semantic index, code graph, and grounded Q&A. TypeScript and Go are next. Editors: VS Code, Cursor, Windsurf, VSCodium, Trae, VS Code Insiders. Agents it drives: Claude Code, Cursor CLI, Codex CLI, Gemini CLI, Copilot Chat, Windsurf Cascade.

Is it really free and open source?

Yes — , every engine builds from source in this repo.

Privacy

• Everything on your machine — per-repo state in `.vinv/` (auto-gitignored), per-machine in `~/.vinv/`. No account, no API keys, no telemetry — none.
• One outbound request, and you can read it: a GET of a static JSON file at `notices.vinv.ai` on activation, for broken-release and security notices only. No query string, no identifiers, nothing uploaded; at most once per 12 hours; disable with `vinv.notices.enabled`.
• The only download is the embedding model (Hugging Face, once, ~100 MB); everything else builds from this repo.
• Traces store bounded summaries, not raw values; sensitive parameter names (`password`, `token`, `api_key`, …) are redacted, never captured.
• The only LLM Vinv talks to is the coding-agent CLI you configured, through its own auth.

Contributing & license

See CONTRIBUTING.md — `uv sync`, `cargo build` in `index/`, `npm install && npm run check` in `extension/`, keep `tests/e2e/planted_bug_golden/run.py` green. Good first issues are labeled. By taking part you agree to our Code of Conduct; to report a vulnerability, see SECURITY.md. Apache License 2.0 © 2026 VinvAI.