---

Why AuraOps

Most “deploy my agent” tools either:

1. force you to write Docker / CUDA / provider SDK glue, or
2. run a demo shell that is not actually your repo.

AuraOps is the middle path:

We optimize for time-to-working-agent and repeatable deploys, not a stopwatch fantasy.

---

What to expect on timing (real)

These are engineering ranges for auraops deploy to Modal — not a guaranteed SLA.

First deploy is almost always the cold path. After the environment and weights are cached, iteration gets much closer to the warm range.

Progress you should feel:

1. Blueprint / package project
2. modal deploy (build or reuse image)
3. Optional weight download into shared volume
4. Live HTTPS endpoint

---

Demo

Demo video coming soon — real project → blueprint → Modal URL, with cold vs warm timing called out (not a fake 4s stopwatch).

---

Quick Start (recommended path)

Local Modal deploy packages your code. You need a Modal account and tokens.

# Optional global CLI
npm install -g auraops
# or: npx auraops ...

# 1) Analyze project → .auraops/blueprint.json
auraops init ./my-agent-project

# 2) Deploy (default = local Modal CLI)
# If MODAL_TOKEN_ID / MODAL_TOKEN_SECRET are missing, the CLI prompts for them.
export MODAL_TOKEN_ID=...
export MODAL_TOKEN_SECRET=...

# Optional: agent API keys injected into the Modal runtime
export OPENAI_API_KEY=...

auraops deploy
# or: auraops deploy -b ./my-agent-project/.auraops/blueprint.json

Hosted API path (packs a code bundle; keep models out of the zip):

export AURAOPS_API_TOKEN=...   # CLI prompts if missing (interactive terminal)
auraops deploy --server --token "$AURAOPS_API_TOKEN"

Server body size: set MAX_REQUEST_BODY_BYTES ≥ 32MB. Bundles auto-skip weight/data files and use a ~25MB code budget after skips.

---

What It Does (accurate)

Works offline / local (no Redis required for init)

• auraops init — Parses requirements.txt, pyproject.toml, conda env files; detects frameworks (PyTorch, LangChain, LangGraph, CrewAI, Transformers, JAX, TensorFlow); estimates GPU tier; writes immutable blueprint.json with checksums.
• Ghost Manifest — If no requirements file, scans Python imports and maps common packages (best-effort; not every PyPI name).
• System dependency hints — Maps packages → apt libs (e.g. opencv → libgl1) and injects them into the Modal image build when present on the blueprint.
• Dependency locking — Best-effort pip-compile → requirements.lock (warns and continues if pip-tools is missing).
• Entrypoint detection — Finds a real entry file; at runtime discovers crew / graph / agent / run / … or falls back to running the script (AURAOPS_INPUT + stdout).

Deploy to GPU (Modal-first)

• Packages your project into the Modal app (add_local_dir → /app), not only a framework scaffold.
• Framework scaffold remains a fallback if user code cannot be loaded.
• Secrets / env — Common keys (OPENAI_API_KEY, etc.) or Modal secrets; optional AURAOPS_AGENT_ENV JSON.
• Weight “fridge” — Remote customModels (S3/HTTP) download into a shared Modal Volume on first miss; later deploys can hit cache.
• Multi-GPU — --gpus 1..8 (Modal-style specs).
• MCP — Optional --mcp exposes tool routes on the same Modal URL; Claude Desktop config printed when enabled.
• CLI credentials — Interactive prompts for Modal tokens and API JWT when missing (non-interactive/CI fails clearly).

Control plane (optional Redis / S3 / hosted API)

• JWT auth, deploy/status/logs/terminate routes
• Redis-backed deployment records and weight metadata registry
• S3 helpers for weight storage
• Provider adapters: Modal (primary user-code path), Lambda Labs (when API key set), AWS / Azure (credential-gated; not full parity with Modal user-code deploy yet)

---

Security & isolation (what is actually true)

• Cloud isolation — Modal (and other clouds) use their container sandboxing. AuraOps does not wrap every cloud job in gVisor.
• gVisor (runsc) — Available when using the local Docker provider and runsc is installed on the host.
• Blueprints — SHA256 fingerprints for environment drift detection; treat them as integrity helpers, not a full supply-chain attestation product.
• Weights & secrets — Prefer Modal secrets for production keys. Project bundles skip .env and large model files; put weights on S3/HF and reference them via blueprint customModels.

---

CLI Commands

Use npx auraops … if the package is not installed globally.

Useful deploy flags:

---

Supported frameworks

---

GPU tiers (Modal selection guide)

Indicative hourly prices used for routing defaults (not a live global market feed). Prices come from provider APIs when available (Lambda Labs instance-types, AWS Pricing API); Modal uses the guide map unless overridden (MODAL_PRICE_T4, AURAOPS_GPU_PRICE_JSON).

---

Architecture (mental model)

Python project
    → init: parse + detect + blueprint (+ lock)
    → deploy: package code + deps image + optional weight volume
    → Modal (primary): live HTTPS endpoint
    → optional: Redis/S3 control plane, other providers (partial)

• Control plane: blueprints, auth, deploy records, weight metadata, API.
• Data plane: your agent process on the GPU provider; weights preferred via S3/HF → provider volume/cache.

Deep dive: docs/architecture/code-map.md · CLI map: docs/cli/cli-code-map.md

---

API (hosted)

Default production API: https://auraops-backend-s2gw.onrender.com (override with AURAOPS_API_URL).

Billing (Razorpay pay-as-you-go)

New accounts get a small free trial (5 credits). Buy more via Razorpay:

# public
curl https://YOUR_API/api/v1/billing/plans

# after login JWT
curl -H "Authorization: Bearer $TOKEN" https://YOUR_API/api/v1/billing/me
curl -X POST -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"packId":"starter"}' https://YOUR_API/api/v1/billing/checkout
# open Razorpay Checkout with response.checkout, then:
curl -X POST -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"razorpay_order_id":"...","razorpay_payment_id":"...","razorpay_signature":"..."}' \
  https://YOUR_API/api/v1/billing/verify

Set on server: RAZORPAY_KEY_ID, RAZORPAY_KEY_SECRET, BILLING_ENFORCE=true.

For the marketing site checkout (https://auraops.vercel.app), set:

CORS_ORIGIN=https://auraops.vercel.app

(The API also allows that origin by default in production CORS.)

---

Requirements

---

Development

npm install
npm run dev           # API server
npm run test:unit     # Fast unit suite
npm test              # Full suite (unit + integration)
npm run type-check
npm run build
npm run lint          # 0 errors required in CI

---

Tech stack

---

Project status (honest)

---

Competitive position (fair comparison)

We do not claim “4s cold start with a 15GB model” as a product guarantee. We claim less glue, real packaging, and honest tradeoffs.

---

Deployment lifecycle

flowchart TD
    UserCode["Python project"] -->|auraops init| Blueprint["blueprint.json\n+ optional lock"]
    Blueprint -->|auraops deploy| Package["Package code\n+ deps image"]
    Package --> Weights["Weight volume\nmiss → download / hit → reuse"]
    Package --> Modal["Modal deploy\nHTTPS endpoint"]
    Weights --> Modal
    Modal --> Live["Agent live"]
    Live -->|idle| Scale["Provider scale-down\n+ optional idle cleanup"]

---

MCP (optional)

auraops deploy --mcp

Exposes MCP-style routes on the Modal URL (e.g. /mcp/tools, /mcp/tools/call). Claude Desktop config can be printed after deploy. Prefer the Modal endpoint for tool calls; stdio wrappers need AURAOPS_AGENT_URL if used separately.

---

CI example

# .github/workflows/deploy.yml
jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 20
      - run: npm install -g auraops
      - run: auraops init .
      - run: auraops deploy
        env:
          MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
          MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
          AURAOPS_NONINTERACTIVE: "1"
          # OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

Use AURAOPS_NONINTERACTIVE=1 in CI so missing secrets fail fast instead of waiting on a prompt.

---

Known limitations

• Cold deploys can take several minutes; AuraOps kills modal deploy after MODAL_DEPLOY_TIMEOUT_MS (default 10 minutes, min 1 min / max 30 min). Raise the env var for heavy first builds, or use pre-warmed images.
• Server bundles are for code (~25MB after skipping models); put large weights on S3/HF + customModels.
• Azure/AWS are not full Modal-parity user-code deploys yet.
• Live cross-cloud price ranking is incomplete.
• gVisor is not applied to every cloud deploy.
• Alpha APIs may change before 1.0.

---

Contributing

PRs welcome. See CONTRIBUTING.md.

Good first issues: good first issue

---

License

Apache 2.0