S+
DashboardGet started

Precision-first code review

Only the comments
worth reading.

Most AI reviewers bury real bugs under false positives. S+ is deterministic-first: it reviews only new lines, proves every finding, maps the blast radius across files, and learns the noise your team waves off. Maximum signal — noise on the floor.

Install free, run locally →Connect GitHub

no API keyno code leaves your machine~60s to first review

SIGNAL / NOISESNR 19.0 dB
signal — real findingsnoise — false positivesclean-as-you-code · diff-scoped

01 — Start free

One line. Then ask your agent to review.

S+ ships as an MCP server your coding agent calls. The agent does the reasoning, S+ supplies the proof. Nothing is billed, nothing leaves your laptop.

$ claude mcp add splus -- npx -y @splus/mcp
# ~/.codex/config.toml
[mcp_servers.splus]
command = "npx"
args = ["-y", "@splus/mcp"]
# any MCP-capable client
command: npx
args:    ["-y", "@splus/mcp"]
▸ then say "review my staged changes with splus" — you’ll get findings with blast radius and a one-tap way to teach away the noise.

02 — The only metric that matters

Precision, not catch-rate.

A reviewer that flags everything is just noise with extra steps. We optimize precision — the share of comments you actually act on — and it climbs as S+ learns each repo. The gap below is your signal-to-noise.

1.0.75.50.250precision 95%false positives 5%reviews over time →
Illustrative. S+ publishes the real per-repo curve in your dashboard — precision is a number we stand behind, not a slogan.
only new lineslegacy code is never re-litigated
0 inferenceon the free local tier — your agent thinks
every findingcites a deterministic anchor, not a vibe
compoundingeach dismissal quiets a whole class

03 — How it works

Deterministic first. The model only judges.

Most of the work happens with zero inference — fast, reproducible, and cheap. The LLM is downstream, and it only ever sees what survived.

  1. 1

    Parse the diff, keep only what’s new

    Tree-sitter builds symbols; clean-as-you-code throws away every line your change didn’t add. No legacy noise, ever.

  2. 2

    Run deterministic collectors

    Secrets, cognitive-complexity deltas, SAST, duplication — each finding anchored to hard evidence: a match, a metric, a graph edge.

  3. 3

    Map the blast radius

    A cross-file call graph resolves who depends on what you touched — at compiler-grade SCIP precision when an index is present.

  4. 4

    Suppress, then judge

    Your repo’s learned filter drops known noise. Only then does an LLM triage what’s left — your agent locally, a frontier model in the cloud.

04 — The moat

It tells you what a change can break.

Line-by-line reviewers see a function in isolation. S+ sees the graph: change a signature and it surfaces every caller across the codebase, flags the ones that cross an API boundary, and tells you how sure it is.

  • PRECISE compiler-grade resolution from a SCIP index — ~97% confidence.
  • HEURISTIC name + import fallback when no index exists — honestly labeled.
  • HONEST every blast-radius claim carries its own confidence.
auth/token.tsvalidateToken()changed
api/login.tshandleLogin()
api/refresh.tsrotate()crosses API
mw/guard.tsrequireAuth()
3 direct callers · 2 files · 1 crosses an API boundary · 97% resolved

05 — Two ways to run

Free on your machine. Paid on your PRs.

Local

Freeforever

For the editor. Review before you commit.

  • MCP server your agent calls — Claude Code, Codex, more
  • Full deterministic engine + blast radius
  • Per-repo learning that quiets your noise
  • No API key · no code leaves your machine
  • Your agent is the judge — zero inference cost
Copy the install line
Team

GitHub bot

Usageper active author

For the PR. Review every change, automatically.

  • Auto-review every PR or on @splus mention
  • Configure auto vs mention, severity gates, path filters
  • Frontier-model judging — only what’s worth your time
  • Team learnings sync across the org
  • Transparent, per-author usage in the dashboard
Connect GitHub →

06 — Why it’s different

Built to be quiet.

CapabilityS+Typical AI reviewers
Reviews only newly-added lines●○
Every finding carries a deterministic anchor●○
Cross-file blast radius with confidence●○
Learns & suppresses your dismissals●◐
Runs fully locally, free, no key●○
Bring your own LLM / provider-neutral●○
Default posturequietloud

07 — Trust

Your code is yours.

  • No training on your code. Ever. Not for us, not for a provider.
  • Configurable retention. Keep diffs for minutes, or not at all.
  • Self-host the engine. The deterministic core runs anywhere.
  • Bring your own LLM. Provider-neutral by design.
  • Published precision. We show our methodology and our numbers.
  • SOC 2 path. Built for the security review from day one.
Read the Trust Center →

Stop reading noise.
Start at the signal.

Install free →Connect GitHub