revio

Agentic code review CLI that combines LangGraph-orchestrated LLM
reasoning with deterministic static analyzers (oxlint / bandit / clippy /
spotbugs / golangci-lint / cppcheck + 30 PLC rules) across
17 language profiles.

$ revio review --commit HEAD
🔍 auth.js ──────────────────────────────────────────────
  → read_file(auth.js:42)
  💭 Line 42 builds SQL via template literal. Where does user_id come from?
  → get_call_sites(getUserById)
  💭 Called from POST /user/:id with req.params.id
  ⚠️ CRITICAL  SQL injection at auth.js:42  (confidence 0.95)
     Evidence: req.params.id → getUserById → `WHERE id = ${id}`
     Counter-considered: ORM auto-escape — ruled out, raw query path

Features

	What
3 modes	review (diff) · audit (full repo) · dedup (find AI redundancy)
17 languages	JS/TS · Python · Rust · Java · Go · C/C++ · Shell · Lua · SQL · Ruby · PHP · Kotlin · Verilog/SystemVerilog · 4 generic · PLC · 8 LLM-only
13 static analyzers	oxlint · bandit · clippy · spotbugs · golangci-lint · cppcheck · shellcheck · luacheck · sqlfluff · rubocop · phpstan · detekt · verilator
Local / self-hosted LLM	Point at any Ollama / vLLM / private endpoint — code never leaves the machine. Free + air-gap + FERPA/GDPR-safe.
PLC support	7 vendor parsers · 30+ PLCopen rules · HW audit · LD/FBD/SFC → ST
RAG	Index your company's coding guidelines, cited inline in findings
Skills	Anthropic Agent Skills spec, dual-layer (project + user)
MCP	Client + server — connects to your tools, exposes its own
dedup --fix + undo	Applies patches AND records snapshot history — revio fix undo reverts any past session
Cross-session memory	"🆕 New since last run: 3" — SQLite-backed
Multi-LLM	Anthropic · DeepSeek · Mistral (EU-sovereign) · OpenAI · OpenRouter · any OpenAI-compatible endpoint
Multilingual REPL	Any human language (en / 中 / de / es / 日 ...) → English findings
Natural-language control	Plain-language commands run reviews and change settings (model / endpoint / key / budget …); out-of-scope asks are flagged at the capability boundary

---

Install

One-click

macOS / Linux — copy-paste into a terminal:

curl -sSL https://raw.githubusercontent.com/witold-andelie/revio/main/scripts/install.sh | bash

Windows — copy-paste into PowerShell:

iwr https://raw.githubusercontent.com/witold-andelie/revio/main/scripts/install.ps1 | iex

The installer walks 7 stages, each with progress visible, and asks before
doing anything that takes more than a few seconds:

1. Verifies Python ≥ 3.11 (offers to install via winget on Windows if missing)
2. Verifies git
3. Asks where to install — defaults to ~/.local/share/revio (macOS / Linux) or %LOCALAPPDATA%\revio (Windows); if you're on a different drive than the default (e.g. you cd D:\tools first on Windows), the prompt offers to install there instead
4. Clones the repo (with git --progress)
5. Installs revio core + Tree-sitter grammars (~150 MB)
6. Asks which optional pieces to install:
   • RAG support (~1 GB: torch + sentence-transformers) — opt-in; skip if you don't index company guidelines
   • Per-language static analyzers — [A] all / [C] pick per language / [N] none. Selecting the languages you actually use significantly improves revio's accuracy on those languages.
7. Adds a revio launcher to ~/.local/bin (macOS / Linux) or %LOCALAPPDATA%\revio\bin on PATH (Windows)

After install, open a new terminal so the updated PATH loads, then cd
into any code folder and run revio — works just like the claude command.

Re-run the same install command anytime to update to the latest main.

Disk footprint

What	Size
Core (agent runtime + CLI + Tree-sitter grammars + 13 profiles)	~150 MB
+ RAG (chromadb + sentence-transformers + torch)	+~1 GB
+ HuggingFace embedding model (first RAG use)	+~80 MB
Per-language analyzer binaries (oxlint / cppcheck / shellcheck / etc.)	~1-30 MB each

A typical install with RAG off and only the user's daily languages
selected sits around 180-250 MB. With everything on it's ~1.5 GB.

Uninstall

# macOS / Linux
curl -sSL https://raw.githubusercontent.com/witold-andelie/revio/main/scripts/uninstall.sh | bash

# Windows
iwr https://raw.githubusercontent.com/witold-andelie/revio/main/scripts/uninstall.ps1 | iex

The uninstaller asks separately whether to also remove your cache
(~/.cache/revio — fix history, findings DB, RAG index) and config
(~/.config/revio — config.toml + custom skills). System-wide analyzers
installed via brew/winget/scoop are NOT touched — they may be useful to
other tools.

Manual install (any OS)

git clone https://github.com/witold-andelie/revio.git
cd revio
python3 -m venv .venv
# macOS / Linux:
.venv/bin/pip install -e ".[js,plc,python,languages]"
# Windows PowerShell:
.venv\Scripts\pip install -e ".[js,plc,python,languages]"

Or straight from GitHub without cloning:

pip install "git+https://github.com/witold-andelie/revio.git#egg=revio[js,plc,python,languages]"

Optional static analyzers (per language)

Language	Tool	macOS / Linux	Windows
JS / TS	oxlint	brew install oxlint / npm i -g oxlint	npm i -g oxlint
Python	bandit	(already in [python] extra)	(already in [python] extra)
C / C++	cppcheck	brew install cppcheck / apt install cppcheck	winget install Cppcheck.Cppcheck
Go	golangci-lint	brew install golangci-lint	winget install golangci-lint.golangci-lint
Rust	clippy	rustup component add clippy	rustup component add clippy
Java	spotbugs	brew install spotbugs (needs JDK)	download from spotbugs.github.io
Shell	shellcheck	brew install shellcheck / apt install shellcheck	winget install koalaman.shellcheck
Lua	luacheck	brew install luacheck / luarocks install luacheck	scoop install luacheck
SQL	sqlfluff	(auto-installed into revio's venv)	(auto-installed into revio's venv)
Ruby	rubocop	gem install rubocop	gem install rubocop (needs Ruby)
PHP	phpstan	composer global require phpstan/phpstan	(same; needs PHP + Composer)
Kotlin	detekt	brew install detekt (needs JDK)	download detekt-cli from GitHub
Verilog / SystemVerilog	verilator	brew install verilator / apt install verilator	scoop install verilator

Missing analyzers don't break anything — revio detects what's installed and falls back to AST + LLM reasoning for the rest.

---

First run

.venv/bin/revio

Triggers a 6-step wizard: pick provider → API URL → key →
model → thinking mode → default profile. Tests the
connection, then saves to ~/.config/revio/config.toml.

---

The 3 modes

review — diff / commit

revio review                           # latest commit
revio review --commit abc1234
revio review --format markdown -o review.md

Tight tool budget. Focus: security + correctness in the changed lines.

audit — full-repo scan

revio audit
revio audit --profile python --budget 30

Heavy on Layer 2 (static analyzers run first). LLM adds semantic context
on top.

dedup — find AI-generated redundancy

revio dedup                            # just report
revio dedup --fix                      # interactive: review & apply each patch
revio dedup --fix --dry-run            # preview patches, don't write
revio dedup --fix --yes                # auto-apply high-confidence (CI)

Detects duplicate functions, single-use wrappers, dead code, repeated
templates — the LLM-generated patterns. --fix refuses to start on a
dirty repo unless --allow-dirty (which stashes first).

Undo any past fix — every --fix session snapshots the affected
files before writing, so you can roll back regardless of git state:

revio fix history                       # list past sessions
revio fix undo                          # restore from most recent session
revio fix undo 2026-05-24T10-15-32.123_a3f9  # restore a specific one
revio fix show <session_id>             # preview what would be reverted
revio fix clean --older-than-days 14    # purge old snapshots

History caps default to 50 sessions / 30 days / 1 MiB per file — tweak
in ~/.config/revio/config.toml under [fix_history]. Works without
git; the git stash path remains as additional safety when present.

---

Configuration

revio config show / edit / path / init

Or directly in ~/.config/revio/config.toml. Examples for the major
providers:

# DeepSeek (cheapest cloud option; default in our wizard)
[llm]
provider = "openai_compat"
api_url = "https://api.deepseek.com"
api_key = "sk-..."
model = "deepseek-v4-pro"

# Mistral (EU-sovereign — recommended for European customers)
# [llm]
# provider = "openai_compat"
# api_url = "https://api.mistral.ai/v1"
# api_key = "..."
# model = "codestral-latest"           # code-specialized 22B, perfect for revio

# Anthropic native
# [llm]
# provider = "anthropic"
# api_url = "https://api.anthropic.com"
# api_key = "sk-ant-..."
# model = "claude-sonnet-4-6"

For per-project overrides, drop a .revio.toml in the repo root —
shadows the user-global config and is meant to be committed.

Switching LLM model / endpoint / key after install

Three paths, easiest first:

Goal	Command (in the REPL)
Browse and pick a model (auto-discovers /v1/models)	/model
List models without picking	/model list (or /models)
Set model directly	/model deepseek-v4-pro
Change endpoint URL	/url https://api.mistral.ai/v1
Rotate API key (masked input)	/key
Open the full config file in $EDITOR	/config
Re-run the 6-step wizard	revio config init

The /model picker hits GET /v1/models on the current endpoint at
runtime, so for any new provider (Mistral, a new Xiaomi API, your
in-house vLLM, etc.) you see the model catalog the provider is
actually serving right now — no need to know model IDs by heart.

Adding more static analyzers after install

The installer asked you to pick analyzers by letter code. To add more
later without re-running the installer:

revio analyzers              # status table — what's installed vs missing
revio analyzers install jcs  # install JS + C/C++ + Shell (same letter codes)
revio analyzers install '*'  # install ALL remaining
revio analyzers menu         # interactive picker

revio detects your OS and uses the right package manager (brew on macOS,
apt on Linux, winget / scoop on Windows). sqlfluff is pip-installed
into revio's own venv. Letter codes are the same as the installer's:

j JS · c C/C++ · g Go · r Rust · a Java · s Shell ·
l Lua · q SQL · v Verilog · u Ruby · h PHP · k Kotlin

---

Local / self-hosted LLM (zero data leaves the box)

revio's openai_compat provider works with any OpenAI-compatible
endpoint — that's the dominant API standard for self-hosted runtimes.
revio doesn't care whether the model is a 4 GB quantized Qwen on a
laptop, a 70 B Llama on a workstation, or a full-power 671 B
DeepSeek-V3 / 405 B Llama-3.1 / 489 B Qwen-Max running on a bank's
own GPU cluster. The same provider config drives all of them.

Compatible runtimes (non-exhaustive): Ollama · vLLM · SGLang ·
llama.cpp server · LM Studio · LocalAI · TGI (HuggingFace Text
Generation Inference) · OpenLLM · Triton Inference Server. If it
exposes /v1/chat/completions, revio works against it.

Open-weight model families that work well with revio (verified):
Mistral / Mixtral / Codestral (EU-sovereign, recommended in Europe) ·
Qwen 2.5 / Qwen 3 (multilingual, esp. strong Chinese) · Llama 3.1 / 3.3 ·
DeepSeek-V3 / R1 · Gemma · Phi-4 · GPT-OSS. Any model with
function-calling support — the agent loop uses tool calls.

Why this matters

Constraint	Why local LLM is the only answer
Student / patient / financial code review	FERPA / HIPAA / GDPR / SOX often forbid sending code to a US API
Banks / insurance / law firms	Internal IP + regulator audit trails — code must stay inside the firewall, often with full-size models running on private GPU clusters
Government / defense / aerospace	Air-gapped by policy; both plc and verilog profiles are valuable here
EU AI sovereignty	French / German / Italian / Czech / Polish customers can run Mistral (EU-headquartered, open-weight) or Mixtral locally — GDPR-compliant by construction
National AI sovereignty	Other jurisdictions with similar mandates (China, Russia, etc.)
Cost at scale	A CS department doing 5 000 audits / semester pays $0 instead of $50-500
Vendor independence	No provider rug-pull / pricing-tier change breaks your CI

One-time setup (any local server)

# ~/.config/revio/config.toml
[llm]
provider = "openai_compat"
api_url = "http://<host>:<port>/v1"   # whatever your local server exposes
api_key = "unused"                     # or your internal token, if any
model = "<whatever-model-id-the-endpoint-serves>"

Then revio audit . — fully offline from this point on. The setup is
model-agnostic: replace the api_url + model line and you're
talking to a different deployment. Examples:

Deployment	api_url	model
Ollama on your laptop	http://localhost:11434/v1	e.g. qwen2.5:7b, llama3.1:8b, mistral:7b, codestral:22b
Mistral cloud (EU-sovereign)	https://api.mistral.ai/v1	codestral-latest, mistral-large-latest, mistral-small-latest
vLLM behind nginx in your DC	https://llm.internal/v1	e.g. mistral-large-2, mixtral-8x22b, deepseek-v3-671b, llama-3.1-405b
Bank's on-prem cluster (full-size frontier model)	https://gpu-cluster.bank.local/v1	whatever the cluster team registers
Air-gapped lab box	http://192.168.x.x:8000/v1	whatever's loaded

/model REPL command auto-discovers whatever models the endpoint
serves by hitting GET /v1/models. No separate config for each model.

/cost shows token counts but silently omits the $ figure for
local models (no cost to misrepresent).

Hybrid setup

Embeddings used by RAG (all-MiniLM-L6-v2, ~80 MB) always run locally
in revio's own process — your indexed guidelines never go through any
API. You can combine local embeddings with a cloud LLM, or go fully
local. Mix and match per .revio.toml.

What works · what's identical · what scales with hardware

Feature	Local LLM (any size)	Cloud LLM
Agent loop · tools · streaming · MCP · --fix	✅ identical	✅
All 13 Layer-2 static analyzers	✅ identical (run as subprocesses)	✅
RAG (embeddings)	✅ always local	✅ embeddings local; LLM cloud
Per-call latency	hardware-bound (laptop: slow · 8×H100: fast)	network-bound
Finding quality on tricky semantic cases	scales with model size — a frontier 671 B on local hardware ≈ frontier cloud	typically high
Cost per audit	$0 (you already paid for the GPUs)	$0.01-$0.30 typical

The point: customers with budget for a serious local deployment
(banks, defense primes, large universities, telcos) get the same
quality as cloud — with full data sovereignty. Customers without that
budget can run a 7-8 B model on a developer laptop and still get
useful static-analyzer coverage + LLM reasoning. Same product, both
extremes of the spectrum.

---

Guidelines (RAG)

Index your team's coding standards so findings cite them directly:

revio guidelines add docs/styleguide.md team-policies/
revio guidelines search "SQL injection prevention"
revio guidelines list / clear / reindex

Supported: .md .txt .rst .adoc .pdf .docx. Index lives at
~/.cache/revio/<repo-hash>/vectorstore/ (per-repo).

During a review:

⚠️ CRITICAL  SQL injection
  Evidence:
    · read_file showed: query = f"SELECT * FROM users WHERE id = {id}"
    · search_guidelines → security_checklist.md / SQL Injection: ...

---

Skills

Markdown files with YAML frontmatter that teach the agent how to handle
specific scenarios. Dual layer:

• .revio/skills/<name>/SKILL.md (project, committable)
• ~/.config/revio/skills/<name>/SKILL.md (user-global)

revio skills list / show <name> / activated

Skills with matching extensions, imports, or filename_patterns
auto-fire; others stay catalog-only until the LLM pulls them via
load_skill(name).

---

MCP — both directions

As a client — consume your existing servers

[mcp.servers.jira]
command = "uvx"
args = ["mcp-server-atlassian-jira"]
env = { ATLASSIAN_TOKEN = "$ATLASSIAN_TOKEN" }

revio connects in parallel at session start, wraps their tools as
mcp_<server>_<tool>, and merges them into the agent's toolkit. Failed
servers degrade gracefully.

As a server — expose revio to other agents

revio mcp-server               # stdio MCP server

Tools exposed: revio_audit, revio_review, revio_dedup,
revio_run_bandit, revio_run_oxlint, revio_run_cppcheck,
revio_search_guidelines, revio_list_profiles, revio_detect_profile.

Register with Claude Code:

{ "mcpServers": { "revio": { "command": "revio", "args": ["mcp-server"] } } }

The server returns patch operations via revio_dedup but never
applies them — the host agent decides what to write.

---

Interactive REPL

revio              # drop into REPL

> review the last 3 commits
> 检查这个项目里有没有重复代码         ← reply comes back in Chinese
> Vérifie src/auth.py pour des fuites  ← reply comes back in French
> clean up the duplicate / junk code   ← runs dedup
> switch the model to claude-opus-4-7  ← changes a setting, no slash needed
> set my api key                       ← prompts securely (key never typed inline)
> what can you do?                     ← lists revio's capabilities
> /model deepseek-v4-pro

Slash command
/help /?	List all commands
/model	Interactive picker (live /v1/models + curated catalog)
/model <name>	Set model directly
/models	List available models on current endpoint
/url /key /config	Endpoint / key / config-file management
/profile <name>	Switch profile
/mode <name>	Default mode for next NL input
/budget <n>	Tool-call budget for this session
/cost	Real token usage + USD cost for the REPL session
/clear /history /exit	Standard

Non-slash input is routed by an intent LLM (multilingual by design) into:

• review / audit / dedup — run the agent in that mode. "Clean up
  the junk / duplicate code" maps to dedup.
• config — change a setting in plain language ("switch the model to
  claude-opus-4-7", "set my api key", "budget 30", "use the endpoint
  https://api.mistral.ai/v1", "show my config", "how much did this cost").
  These are translated to the matching slash command and run through the same
  dispatcher, so anything the slash commands do is reachable by natural
  language too. The API key is never read from the text you type — that
  path always drops into the masked prompt; endpoint changes ask to confirm.
• capability — "what can you do?" prints the capability list.
• out_of_scope — requests beyond revio (write a feature, deploy, run a
  shell command, general questions) are declined with a clear note that
  they're outside revio's capability boundary, plus a reminder of what it
  can do — instead of silently launching a review.

There's a deterministic keyword fallback for when the intent LLM is
unreachable, so settings changes and the boundary message still work offline.

Language: input vs output

revio splits language responsibility along a deliberate boundary:

Layer	Language
Wizard banner · UI labels · slash commands · install scripts · docs · examples	English always (so screenshots / Stack Overflow / support tickets read the same in any locale)
Your natural-language requests	Any language — Chinese, German, French, Spanish, Czech, Japanese, ...
Findings shown to you (title, hypothesis, suggestion, counter-consideration, reflect summary, systemic observations, plan text)	Same language as your request — Chinese in → Chinese out
Tool args (read_file("src/auth.py"), regex patterns) and evidence quotes (verbatim tool output)	English always — for log-greppability and tool compatibility

So you can ask 检查 src/auth.py 看有没有 SQL 注入 and get back finding
cards titled SQL 注入：query 用 f-string 拼接 user_id, with the
suggested fix in Chinese — but the agent's internal read_file() call
still uses the literal path "src/auth.py".

Visual rhythm — owl between tasks

The owl mascot that animates on REPL startup also plays a short (~1 s)
loop after every NL-driven task finishes. Visual separator so the
next prompt feels like a fresh task instead of a continuation. Skipped
on non-TTY (CI). Slash commands (/help, /cost, etc.) don't trigger
it — they're configuration, not "tasks".

---

Profiles

Profile	Layer 1 (AST)	Layer 2 (static)
js	Tree-sitter + symbol graph + call graph + dedup index	oxlint
python	Tree-sitter	bandit
rust	Tree-sitter	clippy
java	Tree-sitter	spotbugs (needs .class)
go	Tree-sitter	golangci-lint
cpp	Tree-sitter	cppcheck
plc	7 vendor parsers + LD/FBD/SFC	30+ PLCopen rules + HW audit
shell	Tree-sitter	shellcheck
lua	Tree-sitter	luacheck
sql	Tree-sitter	sqlfluff (multi-dialect)
ruby	Tree-sitter	rubocop
php	Tree-sitter	phpstan
kotlin	Tree-sitter	detekt (needs JDK)
verilog	Tree-sitter	verilator (--lint-only)
generic	Tree-sitter (Scala / C# / Swift / Julia)	—
LLM-only	MATLAB · R · Verilog · SAS · COBOL · Solidity · Zig · ObjC · Dart	—

Default auto walks the repo, counts file extensions + marker files,
picks the best match. Override with --profile <name>.

---

Token usage & cost

revio reads real usage_metadata off every LLM response. Per call:

  · tokens +1.2k in, +340 out  (Σ 8.4k / 1.9k · 85 tok/s · $0.011)

Session footer:

  session: 6/6 tool calls · 7 findings · 18.1s · deepseek-v4-pro
  tokens:  12.6k in · 573 out · 5 LLM call(s) · avg 32 tok/s · $0.0040

Pricing is fuzzy-matched (DeepSeek / Claude / OpenAI / Mistral / Ollama).
For models we don't have pricing for (new providers, custom endpoints)
the $ figure is silently omitted — token counts and throughput still
show. We never display a misleading $0.00.

/cost in the REPL shows cumulative usage across the whole REPL session.

---

Output formats

revio audit --format stream            # default (TTY)
revio audit --format json -o report.json
revio audit --format markdown -o report.md

JSON / markdown include total_input_tokens, total_output_tokens,
llm_call_count, est_cost_usd on the ReviewReport.

---

Exit codes

Code	Meaning
0	Clean, or only info-severity
1	Operational error (no config, invalid path, ...)
2	At least one CRITICAL finding
130	Ctrl-C

For CI: revio audit . --format json -o report.json and fail the build
on exit 2.

---

What's special about revio

	Other LLM review tools	revio
Static-analysis backbone	One LLM call per file	6 deterministic analyzers + LLM on top
Hallucinated findings	Common (LLM invents paths)	Grounding validator rejects un-read files
Token cost on a large repo	Re-reads whole files	Pulls only enclosing functions via Tree-sitter
Company-specific rules	Generic prompt only	RAG over your own guidelines
PLC / industrial control	Not supported	7 vendor parsers + 30 PLCopen + HW audit
LLM provider	Vendor-locked	Anthropic + any OpenAI-compat
Cross-session memory	Stateless	SQLite history — "🆕 New since last run"
Auto-fix	Text suggestion only	--fix actually edits files
Undo a fix	"Hope you committed before running"	revio fix undo — snapshot-based, multi-step, no git required
Self-hosted / offline	Cloud-only, your code goes to a US API	Any OpenAI-compatible endpoint — Ollama, vLLM, on-prem GPU; FERPA/HIPAA/GDPR-safe

---

Troubleshooting

Wizard didn't run → rm ~/.config/revio/config.toml then revio.

Connection test fails → DeepSeek / Mimo need disable_thinking = true. OpenAI-compat gateways often have their own model IDs.

Findings dropped as "ungrounded" → the validator rejected findings on files the agent never read. Run with REVIO_DEBUG=1 to see the full tool-call trace.

dedup --fix says no patches → the model didn't emit propose_patch calls. Either a strict gateway dropped the tool call, or the model didn't think a mechanical fix was safe (check suggestion in the finding).

Tree-sitter import errors → .venv/bin/pip install -e ".[languages]".

---

License

MIT.