AI coding discipline for OpenAI Codex, GitHub Copilot, and any MCP-compatible agent.
CVE scanning · secret detection · PR gates · audit logs · 61 specialist skills · 10 guard agents · engine scripts.
Andie greets you on first use. ≤2 questions. No bash. No 8-question wizard.
Raven-Codex enforces discipline through three platform-agnostic channels: an MCP server (tool-level checks), a GitHub PR gate (server-side enforcement on every PR), and AGENTS.md (instruction layer your agent reads automatically).
# Clone the engine
bash <(curl -fsSL https://raw.githubusercontent.com/giggsoinc/raven-codex/main/install.sh)Then merge the snippet from config.toml.example into ~/.codex/config.toml:
[mcp_servers.raven]
command = "python3"
args = ["/Users/YOUR_USERNAME/.raven-codex/mcp/server.py"]Restart Codex and ask it to run raven_status. That's it — the raven_* tools are now available.
Copy .github/workflows/raven-pr-gate.yml into your repo and run raven-setup to create .raven/manifest.json. The gate posts a discipline-check status on each PR: CVE scan · secret detection · manifest validation. Make it a required check in branch protection to hard-block merges.
cd YourProject && raven-codex-setupWrites .raven/manifest.json, hook scripts, and the local git pre-commit gate.
Raven classifies every prompt and routes it to the cheapest adequate model:
| Tier | Triggers | Approx cost |
|---|---|---|
| SIMPLE | "fix typo", "rename var", single-file edits | lowest tier |
| MEDIUM | tests, docs, debug, refactor scope | mid tier |
| COMPLEX | architecture, security audit, multi-file reasoning | top tier |
| LOCAL_ONLY | secrets detected in prompt, offline mode | free, on-machine |
- Session token counter + cost shown in banner every session start.
- Stop event writes session summary to
~/RavenVault/sessions/(Obsidian-compatible). - Secrets in your prompt → automatically forced to a local model. The cloud never sees them.
- No telemetry. Local-only. All cost data stays on your machine.
Configure via .raven/.model.env — raven-init writes it for you.
| Component | Count | What it does |
|---|---|---|
| Specialist skills | 61 | Andie · Andie Jr · agent-chaining · ui-router · DB · cloud · security · Oracle (6 specialists) · Salesforce · Odoo · AI/ML · Kafka · K8s · Terraform · FastAPI · log management and more |
| Guard agents | 10 | Always-on discipline — blocks inline SQL, secrets, undeclared stacks, missing architecture |
| Slash commands | 14 | /raven-init /raven-harden /raven-debug /raven-incident /raven-registry-sync /raven-approve /raven-scaffold /raven-search /raven-sync and more |
| Engine scripts | 13 | cve-check · secret-scan · audit-log · emit-violation · db-guard · schema-guard · cve-prompt-guard · pr-gate · obsidian-log · session-start and more |
| MCP server | 1 | raven_status · raven_debug · raven_cve_check · raven_violation · raven_sync_libs |
Token cost is a first-class design constraint. Skills load once on invocation and remain in the context window — smaller skills mean every subsequent message is cheaper.
| Optimisation | Saving |
|---|---|
| Andie (−69% size, 200-word cap, Feynman recap) | −6,852 tok per session |
| db-router pure routing table | −2,560 tok per session |
| ui-router trimmed | −1,564 tok per session |
| agent-chaining trimmed | −1,686 tok per session |
| Obsidian → session-start continuity | ~80 tok of prior context, no cold start |
~57% reduction in skill token footprint vs v2.9.1. In a 20-message session: ~53% fewer context-tokens carried.
raven-codex is the OpenAI Codex / Copilot / multi-platform variant.
giggsoinc/raven is the IDE-native variant.
Both share the same skill set and guard agents. Skills, agents, and engine scripts are kept in sync.
Guardrails before you ship. — Giggso / AntiGravity Projects