SeldonFrame

Build almost any AI agent — from your IDE.

The open-source platform that gives your coding agent the primitives real production agents need: triggers, skills, channels, tools, an owned memory (the Brain), guardrails, evals, one-command deploy, and built-in money rails. Describe an agent in one sentence — minutes later it's answering a real business's phone, chat, and SMS, and booking real appointments. Free to build.


Website · For builders · Docs · Discord

Live demos: HVAC · Med spa · Med spa · Weight loss

---

Ship your first agent in 60 seconds

bash
claude mcp add seldonframe -- npx -y @seldonframe/mcp

Or as an auto-updating Claude Code plugin:

/plugin marketplace add seldonframe/seldonframe
/plugin install seldonframe@seldonframe

> Build me an AI receptionist for an HVAC company in New Orleans.

  ✓ Live at acme-hvac.app.seldonframe.com

No API key. No signup form. Your first workspace is free forever. That one sentence stands up a hosted front office — website, booking page, intake form, CRM — with an AI agent already answering on chat and booking against the real calendar. Add a phone number and it answers calls too.

See a real one — a live workspace, not a mockup:

The generated site — quote intake, reviews, and the chat agent (bottom right)
Its booking page — the real calendar the agent books against

And here's the agent on that site handling a real inbound — checking the live calendar, and, when asked for a price, *refusing to invent one* (that's the quote-guard, not a scripted reply):

A genuine exchange with the chatbot on the live site — nothing scripted:

It reads the real calendar. "Today's fully booked — here are tomorrow's 9/10/11 AM slots."
It won't lie about price. Asked what a visit costs, it says "I don't have exact rates on hand, but I'll have a technician confirm the cost before any work starts" — the quote-guard in action, not a canned line.
It qualifies to book. Asks for the service address to confirm you're in-area, then locks the slot.

Same agent answers on SMS and voice too — the channel is just a config knob.

Click any screenshot — this HVAC workspace is live and books real appointments. More live workspaces: Metro MedSpa · SKINNEY Medspa · Vive Ageless Weight Loss — and more on seldonframe.com.

---

What an agent is made of here

Most "agent frameworks" hand you a chat loop and wish you luck. SeldonFrame gives your IDE agent the full production anatomy:

The agent model — Trigger × Skill × Channel

An agent isn't a chatbot UI; it's three independent axes:

• Trigger — *when it runs*: inbound (a call / chat / email / SMS arrives) · event (a domain event fires — booking.completed, lead.created, invoice.paid…) · schedule (a cron cadence).
• Skill — *what it does*: receptionist · review-requester · speed-to-lead · win-back · digest…
• Channel — *how it speaks*: voice · web chat · SMS · email · internal digest.
• Tools — *what it can touch*: native tools (book against the real calendar · read/write the CRM · send SMS & email · take a message) plus 1,000+ app integrations via Composio (Google Calendar, Sheets, Slack, HubSpot, Notion, …), bound per-agent.

surface: voice | chat (the old receptionist-only knob) is just one point in this space — trigger=inbound. One builder creates any agent; the marketplace sells any agent.

From agent to production *loop*

A production agent is a loop, not a single prompt:

> Trigger → (Model + Tools + State) → Verify → Iterate, bounded by a Stop condition, improved by Evals, kept honest by Observability + Guardrails.

Two non-negotiables drive the roadmap: the checker must be separate from the maker (a model grading its own work is too generous a grader), and the loop must have brakes (or it bills you in silence). Where each primitive stands today:

| Primitive | Status | What's there |
|---|---|---|
| Trigger | ✅ Shipped | Inbound + event triggers on the SeldonEvent bus. booking.completed → review-requester; lead.created → speed-to-lead, both sending outbound SMS/email. |
| State | ✅ Shipped | Agent loop-memory in Brain v2 — agents recall what they did before acting and record after. The review "ask once per customer" throttle is now a memory recall, not a bespoke flag. |
| Verify (maker ≠ checker) | ✅ Shipped | Deterministic validators grade every run — pass rates surface on each agent's health card and /runs; run_agent_evals replays scripted scenarios. Rolling out: the same checker as a hard pre-send gate + an LLM judge for judgment calls. |
| Guardrails / Stop | ✅ Shipped | Quote-guard (never invent prices), enforced read-back before booking, per-contact throttles, booking-policy enforcement (hours · duration · required fields), hard call/iteration caps. Rolling out: generic token-budget brakes for long-looping agents. |
| Generate-by-default | ✅ Shipped | One English sentence → trigger + skill + channel + tool bindings + guardrails, generated together. *"text every customer for a Google review the day after their job — never twice, only if completed"* compiles into a real agent. The generate_agent MCP tool and the Studio wizard both run this pipeline. |
| Record → Agent | 🗺 Roadmap | Skip describing it — screen-share your workflow once. SeldonFrame watches, compiles the steps into a draft agent (skill + tool bindings + eval scenarios derived from the recording), and you deploy in minutes. Live behind a flag at /record; becoming the default on-ramp next. |

The loop, drawn

> One sentence builds it; triggers drive it; the ring is the loop *running* — Model → Tools → State → Verify — with tools orbiting the core, /runs + Brain observing, and the Stop brakes bounding every pass. Everything in the diagram is shipped today. *(Animated SVG — if your viewer freezes it, the still frame reads the same.)*

---
Deeper — the pre-wired stack, the architectural bet, the roadmap: [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md).

Five agents you can ship today

Trigger × Skill × Channel are independent axes — pick one of each and you've architected an agent:

| Agent | Trigger | Skill | Channel | The interesting part |
|---|---|---|---|---|
| AI receptionist | inbound call | qualify + book | voice | Books against the real calendar. Quote-guard, enforced read-back, take_message escalation to the owner. |
| Website chatbot | inbound chat | the same receptionist skill | web embed | Same blueprint, different channel — channels are swappable. |
| Speed-to-lead | lead.created event | first-touch + qualify | SMS | Fires seconds after a form submit, straight off the event bus. |
| Review requester | booking.completed event | review ask | SMS / email | Brain recall enforces "never ask the same customer twice." |
| Anything you describe | any | one English sentence | any | The Studio wizard adapts its questions to the primitives your agent needs. |

The Brain — memory that compounds

Agents that forget are demos. Every SeldonFrame workspace owns a Brain:

• The Soul — the single source of truth an agent grounds on (identity, services, hours, pricing). Install someone else's agent and it re-grounds on *your* Soul, never their facts.
• Loop memory — agents recall before acting and record after. "Did I already ask this customer for a review?" is a memory recall, not a bespoke flag.
• Notes → distilled rules — raw observations get compiled and promoted into durable knowledge on a schedule, so every run starts smarter than the last one ended.

Money rails, built in

Agents that do real work end up touching real money — so the rails are platform primitives, not an integration project:

• A prepaid wallet ledger — UNIQUE idempotency keys and guarded never-negative decrements: a charge can't double-fire, a balance can't go below zero.
• Stripe Connect payouts — seldonframe payout moves accrued earnings to a bank account.
• Checkout that fulfills — a marketplace sale auto-provisions the deployment and re-grounds it on the buyer's business.
• Per-use rental billing — every agent exposes a signed MCP endpoint; other LLMs can rent it and the ledger meters each use.
• Real phone numbers — provision a number per agent (bring your own Twilio today; SF-managed metered numbers are rolling out).

Fee fine print: SeldonFrame takes a 5% GMV fee only when the marketplace brings the buyer, ~2% on sales through your own SeldonFrame storefront, and $0 anywhere else.

---

Use SeldonFrame from any IDE

One npm package — @seldonframe/mcp — runs as a local MCP server in every major AI-native editor. Pick yours, paste the snippet, and ask your agent to build a workspace — your own website, booking page, intake form, and CRM, with the agent wired in. First workspace is free and needs no API key.

Claude Code

bash
claude mcp add seldonframe -- npx -y @seldonframe/mcp

Or as a plugin (auto-updates, one-liner via the marketplace):

/plugin marketplace add seldonframe/seldonframe
/plugin install seldonframe@seldonframe

Cursor

Add to ~/.cursor/mcp.json:

json
{
  "mcpServers": {
    "seldonframe": {
      "command": "npx",
      "args": ["-y", "@seldonframe/mcp"]
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

json
{
  "mcpServers": {
    "seldonframe": {
      "command": "npx",
      "args": ["-y", "@seldonframe/mcp"]
    }
  }
}

VS Code (Copilot agent mode)

Add to .vscode/mcp.json:

json
{
  "servers": {
    "seldonframe": {
      "command": "npx",
      "args": ["-y", "@seldonframe/mcp"]
    }
  }
}

Zed

Add to settings.json:

json
{
  "context_servers": {
    "seldonframe": {
      "source": "custom",
      "command": "npx",
      "args": ["-y", "@seldonframe/mcp"]
    }
  }
}

Codex CLI

Add to ~/.codex/config.toml:

toml
[mcp_servers.seldonframe]
command = "npx"
args = ["-y", "@seldonframe/mcp"]

Or one line: codex mcp add seldonframe -- npx -y @seldonframe/mcp

Once connected, restart your IDE (MCP connectors load at session start), then just say:

> Build a workspace for [business name]. [city, state]. [services]. [phone, optional].

See the same six snippets, kept in sync, at seldonframe.com/build.

---

Self-host the whole thing

Prefer to run it yourself? The entire monorepo is AGPL-3.0 — and there's a prebuilt image, so you don't even build it:

bash
git clone https://github.com/seldonframe/seldonframe.git && cd seldonframe
cp .env.docker.example .env.docker   # add your ANTHROPIC_API_KEY or OPENAI_API_KEY
docker compose -f docker-compose.yml -f docker-compose.ghcr.yml up   # pulls the prebuilt image

Brings up Postgres, migrations, and the app on localhost:3000 — the same dashboard, generated public sites, and API as the hosted version. Multi-arch image (amd64 + arm64) at ghcr.io/seldonframe/seldonframe; or docker compose up --build to build from source. Bring your own LLM key (Anthropic or OpenAI); SMS/voice (Twilio), email (Resend), and 1,000+ integrations (Composio) are add-your-own-key. Full guide: [QUICKSTART.md](QUICKSTART.md#self-host).

---

When you're ready to sell

Your builder hub is seldonframe.com/build — publish an agent, set the price (monthly, per-call, per-outcome, or one-time), and any LLM can rent it over MCP. Every published agent gets a free storefront: a public listing page with schema.org markup, a markdown twin, and llms.txt coverage — so both humans *and* AI agents can discover and buy it. We ship the commodity agents as everyone's free floor, we never build vertical agents that compete with yours, and we never train on your prompts or data.

Pricing — no surprises

| | |
|---|---|
| Self-host | $0 — AGPL-3.0, the entire monorepo · [docker compose up](QUICKSTART.md#self-host) and you're running |
| Hosted — first workspace | Free forever, no card |
| Hosted — unlimited workspaces | $29/mo flat (white-label + voice included) |
| Your AI tokens | Bring your own key — we never mark up usage |
| When you sell | 5% only when the marketplace brings the buyer · ~2% through your own storefront · $0 anywhere else |

Tech stack

• Next.js 16 (App Router) + React 19 + TypeScript — one deployable app: dashboard, generated public sites, and

…

Collapse