Getting Started

Argus runs two ways: as the /argus Claude Code skill, or as standalone Python scripts from a shell. Both share the same config.yaml and environment variables. Pick the track that matches you.

Beginner track

Five steps. One key. You will be reviewing diffs in a couple of minutes.

1. Clone and enter the repo

git clone https://github.com/jimstratus/argus.git
cd argus

2. Install the Python dependencies

pip install pyyaml psutil

3. Point Argus at itself

ARGUS_HOME tells every script where the config and prompts live. Set it to the repo root:

export ARGUS_HOME=$PWD

4. Set one API key

The public default route is OpenRouter — a single key covers most reviewers. Keys live only in your environment; Argus never writes them to disk.

export OPENROUTER_API_KEY="sk-or-..."

5. Generate the aichat config, verify, and review

install_aichat.py writes ~/.config/aichat/config.yaml with the provider client definitions (no keys on disk — keys are forwarded from the environment at dispatch time).

python scripts/install_aichat.py --merge
python scripts/verify.py --all

Then run a review. Inside Claude Code:

/argus

That reviews your current diff with the default standard profile and prints one merged review.

Advanced track

For power users with their own provider subscriptions, large diffs, and benchmark needs.

Multiple provider keys + direct-API routing

Three default-roster reviewers are dual-route: glm-5.2, minimax-m3, and deepseek-v4-pro (the custom-only hermes-4.3 is dual-route too). With their direct keys set, you can prefer each provider’s own API first (cheaper, uses your subscriptions) and keep OpenRouter as the fallback.

export ZAI_API_KEY="..."          # GLM direct (z.ai Coding Plan)
export MINIMAX_API_KEY="..."      # MiniMax direct
export DEEPSEEK_API_KEY="..."     # DeepSeek direct (api.deepseek.com)

# Choose direct-first for this run:
python scripts/verify.py --all --prefer-direct
# or persist it:  set route_preference: direct in config.yaml
# or per-shell:   export ARGUS_ROUTE_PREF=direct

See Routing preference for the full precedence rules. CLI reviewers (Codex / Claude / OpenCode / Gemini) are never reordered — their subscription stays primary.

Profiles

Skip naming individual reviewers — pick a profile:

/argus --profile panel        # maximum coverage
/argus --profile security     # auth / crypto / input focus
/argus --profile deep         # long-context reviewers for large diffs
/argus --profile direct       # direct-API subs only (pair with route_preference: direct)

Manual shell pipeline

RUN_DIR="$ARGUS_HOME/runs/$(date +%Y%m%dT%H%M%S)-manual"; mkdir -p "$RUN_DIR"
git diff HEAD > "$RUN_DIR/diff.patch"
python scripts/estimate_cost.py --roster "glm-5.2,minimax-m3,gemini-or,codex" --diff "$RUN_DIR/diff.patch"
python scripts/dispatch.py --run-dir "$RUN_DIR" --roster "glm-5.2,minimax-m3,gemini-or,codex" --diff "$RUN_DIR/diff.patch"
python scripts/merge.py --run-dir "$RUN_DIR"

Benchmark seeding + parallel shells

Benchmark mode scores reviewers against labeled fixtures to build a leaderboard. Always dry-run a single fixture first to catch provider-config bugs in ~30 seconds:

python scripts/benchmark.py --runs 1 --fixtures sql-injection --profile panel   # smoke test
python scripts/benchmark.py --runs 3 --profile panel                            # full run

The recommended protocol is one shell per reviewer, max five concurrent, with a hard wall-cap per reviewer (--max-wall-sec) so a single stalled provider never blocks the run. Each reviewer writes its own incremental JSON for tailable progress.

Host-CLI awareness

Argus detects the CLI host it runs inside (claude / codex / gemini / opencode / unknown) and never asks a host CLI to review its own invocation. Inside Claude Code the claude reviewer is auto-skipped. For profile rosters, the claude reviewer is auto-added when the host is not Claude. See the host-CLI table.