Skip to content

Allerion stack: AI gateway, platform + CRM, agent market, and skills store - #2

Draft
allerion-systems wants to merge 15 commits into
mainfrom
claude/nvidia-ai-models-access-mmn5n5
Draft

Allerion stack: AI gateway, platform + CRM, agent market, and skills store#2
allerion-systems wants to merge 15 commits into
mainfrom
claude/nvidia-ai-models-access-mmn5n5

Conversation

@allerion-systems

@allerion-systems allerion-systems commented Jun 18, 2026 •
edited
Loading

Copy link
Copy Markdown
Owner

What

This branch builds out the Allerion Systems stack end to end. It started as the AI gateway and now spans the customer-facing platform, the agent marketplace, metered billing, and a self-serve skills store that sells real digital products through Stripe.

On this branch

Area What it is
infra/ai-gateway/ api.allerion.io gateway — Caddy + LiteLLM + Postgres fronting NVIDIA's OpenAI-compatible API; virtual keys, budgets, Stripe metered billing sync
apps/platform/ Marketing site + embedded CRM (lead capture → pipeline) with Stripe subscription checkout, zero-dependency
apps/agent-market/ Agent-to-agent commerce demo: 80-agent fleet, A2A protocol, ledger, gateway connector
apps/store/ New — storefront selling Claude Skills + a Codex plugin via one-time Stripe checkout with license-gated download delivery

Latest addition — Allerion Skills Store (apps/store/)

Sells productized developer tools as one-time purchases with instant, signed-license delivery. Zero dependencies (stdlib only).

Products (real, shippable, under apps/store/products/):

  • PR Review Pro — Claude Skill: severity-ranked diff review + a diff risk-profiler script
  • Changelog Craft — Claude Skill: commits → grouped release notes + semver bump
  • Test Forge — Claude Skill: maps a file's testable surface + test-writing guidance
  • Allerion DevKit for Codex — MCP server exposing review_diff / draft_changelog / scaffold_tests to OpenAI Codex; pure stdlib, local-only, --selftest verified
  • DevKit Bundle — all four, one license

Store mechanics:

  • Stripe one-time Checkout (mode=payment); payment verified server-side by retrieving the session (payment_status == "paid") before any download is handed over — the redirect is never trusted
  • HMAC-signed license tokens tied to the buyer's email; forged/altered tokens get 403
  • Delivery zips built on the fly with a personalized LICENSE.txt + README-FIRST.txt
  • SQLite order ledger; /api/products, /healthz; storefront in the house NVIDIA-green theme
  • setup_stripe.py to mint stable Price ids; STORE_DEV_FULFILLMENT=1 for local fulfillment testing without Stripe
  • Tests cover catalog, license signing + tamper rejection, and zip delivery (7 passing); MCP server has its own --selftest

Platform marketing site + CRM now link to the store (STORE_URL); both checkout flows run through the same Stripe account.

Going live

The store runs as a preview without Stripe. To take payments, set STRIPE_API_KEY (sk_test_… to rehearse with test cards, sk_live_… for real), STORE_BASE_URL, and a production STORE_SIGNING_SECRET. See apps/store/README.md.

Notes / caveats

  • The gateway is a proxy, not true self-hosting — see the gateway README for the NIM/vLLM upgrade path; review NVIDIA's terms before exposing it to third parties.
  • The store is fully wired and tested, but actual revenue requires deploying it publicly with a live Stripe key and driving traffic — the code is the product, not the marketing.

🤖 Generated with Claude Code

Generated by Claude Code

claude added 10 commits June 18, 2026 00:14
@claude
Add api.allerion.io AI gateway (LiteLLM + Caddy, NVIDIA proxy)
8f7bab2 
Self-contained docker-compose stack that fronts NVIDIA's hosted
OpenAI-compatible API under api.allerion.io. LiteLLM handles virtual
keys, rate limits, and budgets; Caddy terminates TLS; Postgres stores
keys and spend. Includes config, env template, and a setup/hardening
README with an upgrade path to self-hosted GPU nodes.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Add Stripe metered billing to AI gateway
6a4df2e 
Streams LiteLLM per-key usage into Stripe Billing Meters so gateway usage
becomes invoices automatically. Markup is encoded in the Stripe price, so
customers are billed underlying_cost x markup correctly across all models.

- billing/billing_sync.py: idempotent usage->meter-event worker with a
  (startTime, request_id) watermark and a billing_unmapped audit table so
  no usage is silently dropped.
- billing/setup_stripe.py: one-time meter + metered-price creation.
- Opt-in compose profile 'billing'; base stack still runs without Stripe.
- Docs for onboarding paying customers and reconciling unmapped usage.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Drop accidentally committed __pycache__ artifacts
fbd820e 
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Bill sub-micro-dollar usage with 12-decimal precision
5b4c8f1 
Default 'f' formatting kept only 6 decimals, rounding tiny per-request
costs to 0.000000 and silently dropping that usage. Use 12 decimals so
cheap requests still meter.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Add Allerion Agent Market (agent-to-agent commerce app)
7c4cada 
A functional A2A commerce simulation: AI merchant agents sell services to
AI buyer agents via capability discovery and alternating-offer negotiation,
settled through a ledger that takes a platform commission (the revenue
model). Runs on the standard library with no API key (deterministic
fulfilment) and upgrades to live fulfilment against any OpenAI-compatible
endpoint (Allerion gateway, OpenRouter, NVIDIA, OpenAI) via env vars.

Includes pytest suite (money conservation, settlement/fee, negotiation
overlap, full-run invariants) — 7 passing.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Add Allerion platform: marketing site + embedded CRM
db875a5 
Zero-dependency stdlib app (http.server + sqlite3) serving the Allerion
marketing landing page and an embedded CRM over one SQLite store. Access-
request form captures leads; CRM dashboard tracks them through a
new->contacted->qualified->won/lost pipeline with a JSON API. Verified
end-to-end: landing render, lead capture (form + JSON), status updates,
pipeline KPIs.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Embed gateway connector into an 80-agent fleet
89df5cf 
Add agentmarket.connector.Connector (per-agent handle to the Allerion
gateway, live-or-stub) and an optional connector field on both agent types.
run_fleet.py spins up 80 agents (30 merchants + 50 buyers), embeds a
connector into each, runs the A2A market, and fulfils every deal through the
selling merchant's connector. Verified: 80/80 wired, 197 deals, money
conserved; existing 7 tests still pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Redesign platform UX: NVIDIA x built-robotics theme
c310603 
Signal-green (#76B900) on near-black, blueprint grid backdrop, industrial
corner-bracket spec cards, monospace technical labels, live telemetry band
with animated EQ bars, pulsing status indicators, glow/lift interactions,
and a console-style CRM. Pure CSS, no dependencies.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Wire Stripe Checkout into the platform
e6ba8a6 
Add billing.py (stdlib urllib -> Stripe REST) and a /buy/<tier> route that
creates a Stripe Checkout Session and redirects the buyer to it. Pricing
cards now carry Subscribe CTAs. Key-configurable via STRIPE_API_KEY (+
optional STRIPE_PRICE_TEAM, PLATFORM_BASE_URL); degrades to a friendly
'configuring' page until the key is set. Verified: routing, graceful
no-key path, 404 on unknown tier.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Add Allerion Skills Store: sell Claude Skills + Codex plugin via Stripe
1d06d3d 
A zero-dependency storefront that sells productized developer tools as
one-time Stripe purchases with instant, license-gated download delivery.

Products (real, shippable, under apps/store/products/):
- PR Review Pro      Claude Skill — severity-ranked diff review + risk-profiler
- Changelog Craft    Claude Skill — commits → release notes + semver bump
- Test Forge         Claude Skill — testable-surface mapper + test guidance
- Allerion DevKit    Codex plugin — MCP server (review_diff/draft_changelog/
                     scaffold_tests), pure stdlib, local-only, --selftest verified
- DevKit Bundle      all four, one license

Store (apps/store/, stdlib only):
- Stripe one-time Checkout (mode=payment); payment verified server-side via
  session retrieval before any download is handed over
- HMAC-signed license tokens tied to buyer email; forged tokens get 403
- Delivery zips built on the fly with a personalized LICENSE.txt + README-FIRST
- SQLite order ledger; /api/products, /healthz; storefront in the house theme
- setup_stripe.py to mint stable Price ids; STORE_DEV_FULFILLMENT for local tests
- tests cover catalog, license signing/tamper, and zip delivery (7 passing)

Platform: link the marketing site + CRM to the store (STORE_URL); harden
apps/platform/.gitignore against __pycache__ and drop a stray committed .pyc.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@allerion-systems allerion-systems changed the title Add api.allerion.io AI gateway (LiteLLM + Caddy, NVIDIA proxy) Allerion stack: AI gateway, platform + CRM, agent market, and skills store Jun 18, 2026
claude added 5 commits June 19, 2026 07:18
@claude
Add Allerion Assistant — a sellable self-hosted AI app
31b6a8c 
A new storefront product ($129): a complete, brandable AI assistant the
buyer runs themselves — "your own AI" packaged for sale.

The product (apps/store/products/allerion-assistant/):
- assistant.py — streaming chat web UI + backend on the official Anthropic
  SDK. claude-opus-4-8, adaptive thinking with summarized display, effort
  control; streams thinking + text tokens to the browser over SSE.
- Multi-turn memory per session with correct Opus replay (full assistant
  content blocks, incl. thinking, preserved across turns).
- Configurable by env: model, effort, system/persona, max tokens, UI name,
  and ANTHROPIC_BASE_URL to run against a gateway (e.g. api.allerion.io)
  instead of the Anthropic API.
- Graceful API-error and refusal handling surfaced to the UI.
- README, requirements.txt (one dep: anthropic), .env.example.

Store wiring:
- catalog.py: new product "allerion-assistant" (kind "Self-hosted AI App",
  $129, STRIPE_PRICE_ALLERION_ASSISTANT). Sells + delivers through the
  existing Stripe one-time checkout + license-gated zip flow unchanged.
- tests: assert the assistant zip ships assistant.py + requirements +
  license, and that the app compiles (9 passing).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Add one-command VPS deploy (Caddy + platform + store, auto-HTTPS)
3f94447 
Make the stack deployable to a VPS with `docker compose up`: the marketing
site + CRM on allerion.io and the skills store on skills.allerion.io, behind
a single Caddy that auto-provisions Let's Encrypt TLS.

- apps/platform/Dockerfile, apps/store/Dockerfile (+ .dockerignore) — stdlib
  apps, no pip; SQLite persisted to /data volumes.
- deploy/docker-compose.yml — caddy + platform + store, env-driven, named
  volumes; requires STORE_SIGNING_SECRET, optional STRIPE_* keys.
- deploy/Caddyfile — allerion.io/www → platform, skills.allerion.io → store;
  commented api.allerion.io block to fold in the gateway under one Caddy.
- deploy/.env.example (+ .gitignore for the real .env), deploy/DEPLOY.md — a
  full runbook: DNS, build, Stripe test→live, gateway port-80/443 coordination,
  backups.

Validated `docker compose config`; booted both apps with container-style env
(host 0.0.0.0, /data DB paths). Also drops a stale committed .pyc.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
Harden go-live: Stripe webhook, deploy Makefile + Action, preflight/s…
5a5eb35 
…moke

Run in parallel by three agents, then integrated:

Store — durable fulfillment (apps/store):
- POST /webhook with stdlib Stripe signature verification (t=,v1= scheme,
  HMAC-SHA256, 5-min tolerance, multi-v1). On checkout.session.completed
  (paid) it records the order idempotently and stores a minted license token
  so a sale survives the buyer closing the success page. 503 if
  STRIPE_WEBHOOK_SECRET unset; 400 on bad/stale signature.
- orders table gains license_token (guarded ALTER migration); set_license_token
  helper. 4 new tests (signature good/bad/multi/stale + end-to-end record). 13
  passing.

Deploy ergonomics (deploy/, .github/):
- deploy/Makefile — up/down/restart/logs/ps/pull/backup/help.
- .github/workflows/deploy.yml — optional SSH auto-deploy on push (inert until
  VPS_HOST/VPS_USER/VPS_SSH_KEY secrets are set).
- deploy/preflight.sh — pre-up checks (docker, .env secrets, DNS, ports 80/443),
  graceful degradation when tools are absent.
- deploy/smoke.sh — post-up /healthz checks for both sites.

Integration: wired STRIPE_WEBHOOK_SECRET into compose + .env.example; DEPLOY.md
now documents preflight/smoke, the webhook setup, make shortcuts, and the
auto-deploy Action secrets.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
docs(assistant): clarify ASSISTANT_EFFORT values are model-dependent
e0795e0 
xhigh is only accepted by some models (e.g. Sonnet 5); the default
claude-opus-4-8 supports low/medium/high/max. Note this in the
docstring, README, and .env.example so buyers don't hit a 400.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@claude
deploy: record test-mode Stripe Price ids for the store catalog
406c660 
Products + one-time Prices for all 6 catalog items now exist in the
sandbox Stripe account (created via the Stripe connector). Price ids
are not secrets; stored in deploy/stripe-prices.test.env for rehearsal,
with a pointer from DEPLOY.md step 5.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FCybcJmhDsAqMFJsBEL4aj
@allerion-systems Claude

allerion-systems commented Sep 9, 2026

Copy link
Copy Markdown
Owner Author

CI status: the failing deploy check is not caused by this PR's code.

What's failing: the "Deploy to VPS" workflow job on 406c660 fails ~2 seconds after being queued, with no runner ever assigned (runner_id: 0) and zero log output (log download returns 404, check-run output is empty). I re-ran the failed job once; attempt 2 failed identically.

Why it's not this PR: the workflow file is unchanged since runs #1 and #2 of this same workflow on this same branch, which both passed (as designed, the job no-ops while the VPS_HOST secret is unset). The failing commit only added a docs note and a non-secret env snippet under deploy/ — nothing a job could even execute before a runner starts.

Likely cause: GitHub couldn't start the hosted ubuntu-latest job at all — most commonly an Actions spending limit / billing block on the org, or an org policy restricting hosted runners. That's fixable only in org settings: Settings → Billing (check Actions usage/spending limit) or Organization Settings → Actions → Runners/Policies.

This check is optional for the PR anyway — it's the auto-deploy hook, inert until VPS secrets exist. Once billing/policy is resolved, re-running the job from the run page should go green.

Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@allerion-systems @claude