acyl-app-sec

Personal / local AppSec platform — scan your own repositories on your machine (air-gapped OK), triage findings with a Foundry-lite harness, and open reviewable fix PRs.

Inspired by the Aikido-style scan → denoise → fix loop, but fully local:

Quick start

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Deterministic scan of the intentional fixture (no GPU required)
ACYL_MODEL_MOCK=1 acyl serve-model --mock --port 8080 &
acyl scan fixtures/vulnerable-app --no-docker
acyl findings <run_id> --state confirmed
acyl report <run_id> --open

Web dashboard

acyl dashboard
# open http://127.0.0.1:8888

The dashboard lists runs, shows confirmed / needs-review findings with evidence, renders reports, and can queue new scans or offline autofixes. It binds to localhost by default.

Docker (recommended for “download and run”)

# From a clone:
docker compose up --build
# open http://127.0.0.1:8888  → scan path: /targets/app

# Your own repo:
ACYL_TARGET=/absolute/path/to/repo docker compose up --build

# Or one-shot:
docker build -t acyl:local .
docker run --rm -p 127.0.0.1:8888:8888 \
  -v acyl-data:/data \
  -v /absolute/path/to/repo:/targets/app:ro \
  acyl:local

Published images (after merge to main): ghcr.io/avinn1990/acyl-app-sec:latest
Full guide: docs/DOCKER.md.

Scan any local checkout (public or private). No per-repo goals.md required — acyl uses bundled goals/standard.md by default:

acyl scan /path/to/repo --no-antares
# prints: Using goals: …/goals/standard.md

Override with --goals, a repo-local goals.md, or ACYL_GOALS_FILE. See docs/GOALS.md.

Private repos via git URL use your existing host credentials (ssh-agent, gh auth, git credential helper). acyl does not store PATs.

acyl scan --git-url git@github.com:org/private-repo.git --no-antares

Local install from scratch

Full step-by-step for a fresh machine (macOS or Linux), including real Antares inference, dashboard, everyday workflow, Docker-only option, Ollama permanent hosting, and troubleshooting:

→ docs/LOCAL.md

Antares model hosting

Cisco Antares-350M (fdtn-ai/antares-350m) is a gated, on-prem vulnerability-localization model built on IBM Granite 4.0 350M. acyl runs it only on the operator machine via an OpenAI-compatible localhost server; the Antares agent loop (outside the model) executes allowlisted shell commands in a Docker sandbox with network=none (or an in-process allowlisted shell when Docker is disabled).

Autofix uses a separate optional endpoint ACYL_FIX_MODEL_URL. SCA bumps and secret redaction stubs work with no LLM. For CI / dogfood without weights: acyl serve-model --mock.

For always-on shared inference callable by any local OpenAI-compatible web app, see docs/LOCAL.md § Permanent shared inference (Ollama).

Install Antares (local / Mac Mini)

Validated on Apple Silicon Mac Mini (8GB+). Python 3.12+, Git, and optional Docker Desktop.

1. Request gated access to fdtn-ai/antares-350m on Hugging Face (manual review).
2. Clone and install model extras:

git clone https://github.com/avinn1990/acyl-app-sec.git
cd acyl-app-sec
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[model]"   # pulls torch, transformers, accelerate

3. Authenticate to Hugging Face (do not put tokens in source, LaunchAgents, or compose files):

huggingface-cli login
# token is stored under ~/.cache/huggingface/ (or ~/.huggingface/)
# optional macOS Keychain:
#   security add-generic-password -a "$USER" -s huggingface -w 'YOUR_TOKEN' -U

4. Start the server (first run downloads weights into ~/.cache/acyl/models/):

acyl serve-model --host 127.0.0.1 --port 8080

5. Verify health and real generation (health alone is not enough — see troubleshooting):

curl -s http://127.0.0.1:8080/health
# → {"ok": true, "mock": false, "model": "fdtn-ai/antares-350m"}

curl -sS http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"fdtn-ai/antares-350m","messages":[{"role":"user","content":"Say hello in one word."}],"max_tokens":64}'

6. Point acyl at the server (defaults already match localhost:8080):

export ACYL_MODEL_URL=http://127.0.0.1:8080/v1   # optional; this is the default
export ACYL_MODEL_ID=fdtn-ai/antares-350m          # optional; this is the default
acyl scan /path/to/repo          # omit --no-antares
acyl dashboard                   # UI: Antares on by default; check “Skip Antares” to disable

The web UI does not call Antares directly and has no inference-URL field. The browser talks to the dashboard API; the backend ChatClient uses ACYL_MODEL_URL. To use a different OpenAI-compatible endpoint, set ACYL_MODEL_URL / ACYL_MODEL_ID in the environment that runs acyl (CLI or dashboard).

Keep it running on macOS (launchd)

Foreground serve-model works for smoke tests. For always-on hosting after login, use a LaunchAgent that calls the absolute venv binary (launchd does not load .zshrc / PATH):

mkdir -p "$HOME/bin" "$HOME/Library/Logs/acyl"
cat > "$HOME/bin/acyl-serve-antares.sh" <<'EOF'
#!/bin/bash
set -euo pipefail
export HOME="${HOME:-/Users/REPLACE_ME}"
export PATH="/opt/homebrew/bin:/usr/local/bin:/usr/bin:/bin"
# Edit REPO to your clone path (example: /Users/you/acyl-app-sec)
REPO="$HOME/acyl-app-sec"
exec "$REPO/.venv/bin/acyl" serve-model --host 127.0.0.1 --port 8080
EOF
chmod +x "$HOME/bin/acyl-serve-antares.sh"

Create ~/Library/LaunchAgents/ai.acyl.antares.plist with absolute paths only (no ~), RunAtLoad, KeepAlive, and HOME in EnvironmentVariables. Log to ~/Library/Logs/acyl/antares.{out,err}.log. Then:

launchctl bootstrap "gui/$(id -u)" ~/Library/LaunchAgents/ai.acyl.antares.plist
launchctl kickstart -k "gui/$(id -u)/ai.acyl.antares"

After pulling code fixes:

cd ~/acyl-app-sec && git pull && source .venv/bin/activate && pip install -e ".[model]"
launchctl kickstart -k "gui/$(id -u)/ai.acyl.antares"

Keep the Mini awake when plugged in. Prefer Ethernet. FileVault still requires unlock after power loss for a user LaunchAgent.

Findings from local install

Troubleshooting

Useful probes:

launchctl print "gui/$(id -u)/ai.acyl.antares"
tail -n 100 ~/Library/Logs/acyl/antares.err.log
python -c "import transformers; print(transformers.__version__)"
ls -lah ~/.cache/acyl/models/

Foundry-lite harness

Roles in v1: Orchestrator, Indexer, Cartographer, Detector (CodeGuard + secrets + SCA + Antares), Triager, Reporter, gated Autofix. Validator / exploited is disabled until a disposable testbed exists.

Scans run as a multi-agent task queue over SQLite (tasks / claims with heartbeat reclaim). Worker caps: 1 for indexer, cartographer, triager, and reporter; 2 for the shared detector pool (secrets / SCA / CodeGuard / Antares goals).

See specs/001-foundry/ for the constitution and clarifications.

Cursor: CodeGuard agents & skills

This repo vendors Project CodeGuard v1.4.0 under .cursor/ so Cursor picks it up from any clone (no global install):

Attribution: .cursor/CODEGUARD_NOTICE.md (CC-BY-4.0). Re-sync with scripts/vendor-codeguard.sh.

These Cursor assets are separate from the runtime detector rules in rules/ used by acyl scan.

Air gap

• Cache Antares weights under ~/.cache/acyl/models/
• Vendor CodeGuard rules ship in-repo (rules/, vendor/codeguard/, .cursor/)
• Secrets/SCA fallbacks work without downloading scanners
• acyl fix --finding <id> --offline writes a patch + fix-branch.sh without pushing

CLI

acyl scan [PATH] [--git-url URL] [--goals FILE] [--no-antares] [--llm-codeguard]
acyl status RUN_ID
acyl findings RUN_ID [--state confirmed]
acyl report RUN_ID [--open]
acyl coverage RUN_ID
acyl fix --finding FIND_ID [--run RUN_ID] [--offline]
acyl serve-model [--mock] [--port 8080]
acyl dashboard [--host 127.0.0.1] [--port 8888]

License

MIT for acyl code. Project CodeGuard materials under vendor/codeguard/, rules/, and .cursor/ are CC-BY-4.0 (see vendor/codeguard/NOTICE.md and .cursor/CODEGUARD_NOTICE.md). Antares weights are Apache-2.0 from Cisco Foundation AI / Hugging Face.