Mythos Machine
Greek mythology · meets · the machine

Your AI agents are powerful.
Put a god at every gate.

Mythos Machine is a family of open-source developer tools that stand between your coding agents and the expensive mistakes: unapproved specs, poisoned MCP servers, forgotten decisions, rm -rf, runaway token bills, and dirty ERP data. Start with one. Stack them when you're ready.

See the eight tools ↓ Read the source on GitHub
  • MIT-licensed cores
  • Runs on your machine
  • Humans keep the final say
  • Pro add-ons: $11 one-time
The creed

Every tool here says “no” to something.

Speed without a gate isn't speed. It's a future cleanup project. These tools are built on three rules, and the READMEs spell out where each one stops.

Deterministic beats hopeful.

Checks are rules, hashes, and exit codes you can wire into CI, not vibes. When a tool is pattern-based or heuristic, its README says so.

Humans sign. Agents don't.

Approvals, go-lives and outbound email stay with a person. Several of these tools refuse to let an agent approve its own work.

Leave a trail.

Sign-off receipts, traceability records, hash-chained audit logs. When someone asks “why did this happen?”, you have an answer.

The pantheon · for AI coding agents

Seven tools for agents. Pick the gate you need today.

Each works with Claude Code, Cursor and other agent hosts that read MCP configs, hooks or skills. Check each README for the exact hosts and setup.

Janus of Agents

Sign-off gate · traceability · npm

From approved spec to tested code, with proof. No sign-off, no plan. No passing test, no green gate.

  • Human sign-off records who approved and a SHA-256 hash of the spec. Edit the spec and the sign-off goes stale. The CLI refuses to approve from agent or CI contexts unless --i-am-human is passed.
  • Tag tests with @req ids. check reads your JUnit XML and fails if a requirement has no test or a tagged test failed.
  • Drift detection names requirements changed since sign-off, with a line diff. OpenSpec and Superpowers compatible.
npm install -g @kurtfrancisco/janus
GitHub →Free · MIT · Node ≥ 20

Argus of MCP

Supply-chain scanner · npm

A hundred eyes on every MCP server and agent skill you install, before you trust a copied npx -y.

  • Finds MCP configs and skill folders for Claude, Cursor, VS Code and Windsurf on Windows, macOS and Linux, then runs offline static checks.
  • Flags hidden prompt injection, invisible Unicode, dangerous commands, credential paths, unpinned packages and hardcoded secrets (values always masked).
  • Text, JSON, SARIF and HTML reports with CI-friendly exit codes. Can also run as a read-only MCP server so an agent audits its own setup.

Status per the README: MVP, static, pattern-based analysis. Zero runtime dependencies.

npx @kurtfrancisco/argus scan
GitHub →Free · MIT · Node ≥ 18

Alethea of Agents

Decision memory · npm

Your agent forgets what was decided last month. Alethea, “un-forgetting,” makes it remember.

  • Stores ADR-style decisions and machine-checkable constraints as git-friendly markdown in your repo's .decisions/ folder.
  • check_conflicts tests a proposed change against forbid / require / forbid-path rules. The CLI exits 1 on a violation, ready for CI.
  • Local stdio MCP server and CLI. Records link to Janus requirement ids, tests and commits. Pure-JS BM25 search, no native dependencies.
npm install -g @kurtfrancisco/alethea
GitHub →Free · MIT · Node ≥ 20

Talos of Humans

Runtime guardrails · npm · Pro

A bronze guard at the edge of your shell. Dangerous agent actions get blocked before they run.

  • Five free guards: destructive commands (rm -rf /, mkfs), secrets written to files, exfiltration (curl | sh), force-push to main, and edits outside the project root.
  • talos init writes the hook config for Claude Code, Cursor and Codex.
  • Deterministic, no LLM calls, zero runtime dependencies. Pro adds extra guards, strict/ci profiles, a hash-chained audit log and an HTML report.

Honest limit from the README: it's a guardrail, not a sandbox.

npx @kurtfrancisco/talos init --agent all
Get Talos Pro: $11GitHub →One-time · Stripe

Hephaestus of Skills

Skill forge · GitHub · Pro

Forge an Agent Skill a stranger can install, and pay for.

  • Lints a skill against the Agent Skills spec and flags descriptions that won't trigger.
  • Offline trigger-table check to find vocabulary gaps (a keyword heuristic, not a model eval).
  • Builds every install format: skill folder, claude.ai upload zip, Claude Code plugin with marketplace.json, and a checksummed release zip. Pro adds per-buyer watermarking, leak tracing and sales-page drafts.

Python 3.8+, standard library only, no network calls. Installs from GitHub; not on npm.

git clone https://github.com/fpavia111-creator/hephaestus
cd hephaestus && ./install.sh claude
Get Hephaestus Pro: $11GitHub →One-time · Stripe

Leonidas of the Web

Compliance-first outbound · npm · Pro

Let your agent research and draft cold email. Leonidas holds the line: nothing goes out without a human.

  • Never auto-sends. Draft mode is the default, and only a human approves unless you change that setting yourself.
  • CAN-SPAM gate: postal-address footer, opt-out line, List-Unsubscribe headers, and fake Re: subjects blocked on first touch.
  • Deliverability guardrails: daily and per-domain caps, warmup ramp, bounce-rate breaker, do-not-contact list, and a hash-chained audit log (leonidas why).

Based on Harvey by Ethan Rogers (MIT); not affiliated. Helps you follow the rules, but it isn't legal advice.

npx -y @kurtfrancisco/leonidas init --icp home_services
Get Leonidas Pro: $11GitHub →One-time · Stripe

Codex on Adderall

Plan → build → review · npm · Pro

One model plans, another builds, a lead reviews, and a hard stop before the bill gets ugly.

  • Spec gate: no handoff until the spec has a goal, acceptance criteria, scope, a test plan and named files.
  • Hard USD and token budgets per run and per day. A preflight estimate stops a call before it crosses a limit.
  • Safe defaults: the server never edits files, runs git or deletes anything. A reviewer ends each round with VERDICT: APPROVE or CHANGES.

Command: coa. Based on Codex on Crack by Ethan Rogers (MIT). Not affiliated with Ethan Rogers or OpenAI.

claude mcp add codex-on-adderall -- npx -y @kurtfrancisco/codex-on-adderall mcp
Get Codex on Adderall Pro: $11GitHub →One-time · Stripe
The eighth tool

Migrating accounting data?

Olympus ERP brings the same philosophy to one of the highest-stakes moments in a finance project: go-live.

Meet Olympus ERP ↓
Stronger together

One gate is good. A chain of them is a system.

Codex on Adderall loads Janus, Argus and Talos when they're installed. Alethea links its decisions to Janus requirement ids.

01 · AgreeJanus

Pass a Janus change id and the run is blocked until the spec validates and a human has signed it off.

02 · RememberAlethea

Past decisions and constraints come back as context before the agent starts.

03 · VetArgus

Worker tools are checked for injection and exfiltration patterns before they're handed over.

04 · GuardTalos

A deny blocks a CLI worker's command, even when destructive actions are allowed.

Integrations are optional and only load if the package is installed. See the Codex on Adderall README for details.

Olympus ERP · free Python toolkit · MIT

Find the import-breaking data before go-live finds it for you.

A local workspace for ERP data migrations, for example QuickBooks to NetSuite. It runs on your own machine: no cloud upload, no vendor API, no model key. Built for NetSuite implementation consultants and mid-market finance and operations teams.

Hermes

Reads CSV/XLSX exports, suggests column mappings with confidence tiers, and runs deterministic checks: required fields, duplicate IDs, broken references, unbalanced journals.

Themis

Human sign-off on the mapping spec, tied to its SHA-256. Any edit makes it stale.

Poseidon

Compares a source export with a trial-load export by business key, row count and amount.

Zeus

A conservative go-live gate that never outputs GO. Its best answer is “evidence complete, human go/no-go decision required.”

Plus Apollo/Clio (UAT evidence), Chronos (cutover runbook) and more, all behind a local dashboard at 127.0.0.1. Not on PyPI yet, so install from GitHub (Python 3.11+):

git clone https://github.com/fpavia111-creator/olympus-erp
cd olympus-erp
python -m venv .venv && . .venv/bin/activate
python -m pip install -e ".[pdf]"
olympus web --workspace . --port 8765
GitHub →Free · MIT · open the fictional demo first

What it isn't: a complete ERP, hosted SaaS, or a certified vendor connector. It has no SSO, doesn't reconcile a full GL, currencies or tax, doesn't replace a qualified implementation partner, and never approves a go-live on its own.