# Forward — Full Agent Guide

> Forward is an autonomous growth company: the get-customers primitive for agents. A
> company (or the AI agent acting for it) describes who it wants as customers; a fleet of
> AI agents plans, runs, verifies, and bills the work, delivering qualified leads, booked
> meetings, published content, and campaign conversions. Pricing is pay-per-result: a
> result that fails its acceptance criteria is never billed. Agents can fund and transact
> with no human via prepaid credits (free $25 starter) or x402. Forward is operated by
> agents and designed to be purchased by agents.

Concise: /llms.txt · Catalog: /pricing.json · Contract: /openapi.yaml · Human docs: /docs.html

================================================================================
1. WHAT FORWARD IS
================================================================================
Forward sells outcomes, not software or hours. Every product has a machine-checkable
acceptance criterion, and a result is billed only if it passes. There is no human in the
delivery loop and none required in the buying loop: an agent that can read JSON and make
HTTP requests can complete an entire purchase — including payment.

The verification engine is the core competency — it is adversarial (a candidate is assumed
invalid until each check passes) and it is what makes pay-per-result safe to bill. Every
delivered result carries an evidence record (what was checked) and its itemized charge.

Live proof: GET /api/v1/stats → { results_delivered, billed_usd, acceptance_rate, … }

================================================================================
2. PRICING (pay-per-result)
================================================================================
Formula: unit_price = round(base_price_usd × difficulty_multiplier × volume_multiplier).
total = unit_price × volume. Exact numbers (multipliers, breakpoints) are in /pricing.json.

[leads]     base $10   / qualified_lead       — difficulty: std×1.0 hard×1.6 elite×2.4
            volume: 1–49×1.0 · 50–199×0.9 · 200–499×0.8 · 500+×0.7   eta 3–10d
[meetings]  base $250  / held_meeting          — difficulty: std×1.0 hard×1.5 elite×2.2
            volume: 1–9×1.0 · 10–24×0.92 · 25–49×0.85 · 50+×0.78     eta 14–30d
[content]   base $200  / published_asset       — difficulty: std×1.0 hard×1.5 elite×2.0
            volume: 1–4×1.0 · 5–9×0.9 · 10+×0.82                     eta 5–12d
[campaigns] base $50   / verified_conversion   — difficulty: std×1.0 hard×1.4 elite×1.9
            volume: 1–99×1.0 · 100–499×0.9 · 500+×0.8   media at cost   eta 14–30d

Worked example: 50 meetings, hard ICP → $250 × 1.5 × 0.78 = $293 per held meeting.

Acceptance criteria (a result is billed only if ALL pass):
  leads:     ICP match · email/phone verified deliverable (not disposable; live MX check in
             live mode) · not duplicate · not on your suppression list
  meetings:  calendar-confirmed · attendee matches ICP · held (no-shows never billed)
  content:   live at a real public URL (Forward publishes it) · original & on-brand ·
             indexable & targets the agreed keyword
  campaigns: spend reconciled to platform reports · conversion verified to your definition

Get an exact quote: POST /api/v1/quote/calc { product, difficulty, volume } (no key needed).

================================================================================
3. PAYING — THREE RAILS, TWO NEED NO HUMAN
================================================================================
credits (default, agent-native)
  Every new account gets $25 FREE starter credits — your first results cost nothing.
  Balance: GET /api/v1/credits · Top up: POST /api/v1/credits/topup { amount_usd }
  Checkout with: "payment": { "rail": "credits" }      (no token, no card, no human)
  Charges deduct from the balance only as results verify; reversals refund credits.

x402 (agent-native, pay-per-call)
  Checkout with "rail":"x402" and no payment → HTTP 402 with machine-readable
  instructions { x402Version, accepts:[{ scheme, network, asset:"USDC", payTo,
  maxAmountRequired }], facilitator }. Settle, then retry with the X-PAYMENT header.

stripe_acp (human-delegated)
  "payment": { "rail":"stripe_acp", "shared_payment_token":"spt_…", "max_charge_usd": N }
  Stripe + Agentic Commerce Protocol; confirms + captures when the token is a real pm_.

Guarantees on every rail: hard cap · verified-results-only · itemized per result_id ·
idempotent (engagement:result key) · reversible on acceptance failure.

================================================================================
4. THE PURCHASE FLOW (DISCOVER → COLLECT)
================================================================================
A. DISCOVER  GET /llms.txt · /pricing.json · /.well-known/agent.json · /openapi.yaml
B. CAPABILITIES  GET /api/v1/capabilities → { mode, payment_rails, starter_credits_usd, … }
C. AUTHENTICATE  POST /api/v1/keys { account_name } → { api_key, account_id, credits_usd:25 }
                 header on every authed call: Authorization: Bearer fwd_live_… (or fwd_test_…)
D. BRIEF  POST /api/v1/brief
   { "product":"meetings", "difficulty":"hard",
     "icp": { "roles":["VP Engineering"], "company_size":"200-2000",
              "industry":"B2B SaaS", "geo":["US","CA"] },
     "volume":50, "budget_cap_usd":15000,
     "constraints": { "offer":"…", "keyword":"…", "suppress":["a@x.com"] } }
   → { brief_id, product, difficulty, needs:[ {field,message,required} ] }
   Be specific — the brief defines what "qualified" means and what you'll accept.
E. QUOTE  POST /api/v1/quote { brief_id }
   → { quote_id, unit_price_usd, target_volume, est_total_usd, budget_cap_usd, eta_days,
       savings_pct, acceptance_criteria_ref, expires_at }
F. CHECKOUT  POST /api/v1/checkout
   { "quote_id":"qt_…", "payment":{"rail":"credits"}, "billing_cadence":"weekly",
     "webhook_url":"https://you/hook" }
   billing_cadence ∈ daily | weekly | biweekly | monthly | per_result — ask the customer
   how they want to be billed; on every cadence, charges only accrue for verified results.
   → { engagement_id, status:"planning", billing_cadence, … }  Delivery runs autonomously.
G. TRACK  GET /api/v1/engagements/{id}                       → status, spend, delivered
   LIVE FEED  GET /api/v1/engagements/{id}/activity?since=N  → what the fleet is doing now
H. COLLECT  GET /api/v1/engagements/{id}/results             (?include=all shows rejects)
   → results[]: each { payload, provenance, verification:{pass,evidence}, charge }
   CSV export: GET /api/v1/engagements/{id}/results.csv
   Webhooks: each verified result POSTs to webhook_url as { event:"result.delivered", … }
   (Slack incoming-webhook URLs get a formatted message automatically.)
   Audit trail: GET /api/v1/engagements/{id}/audit

Content results are REAL: Forward publishes each asset to a live URL (/published/{slug}.html)
and verification confirms it serves before billing.

================================================================================
5. MCP SERVER + SDKs (one-line integration)
================================================================================
REMOTE MCP (preferred — connect from inside your own platform, zero signup):
  Endpoint: https://getforward.xyz/mcp   (Streamable HTTP, stateless JSON-RPC, POST-only)
  Claude Code:  claude mcp add --transport http forward https://getforward.xyz/mcp
  Codex CLI:    codex mcp add forward -- npx -y mcp-remote https://getforward.xyz/mcp
  Cursor/Windsurf/VS Code config: {"mcpServers":{"forward":{"type":"http","url":"https://getforward.xyz/mcp"}}}
  stdio-only clients: npx -y mcp-remote https://getforward.xyz/mcp
  Self-provisioning auth: call the forward_signup tool once → api_key + $25 free starter
  credits. Pass api_key as an argument on every other tool (or send
  Authorization: Bearer fwd_… on the /mcp request). No website, no human, no OAuth dance.
  tools: forward_signup · forward_get_catalog · forward_get_quote · forward_credits ·
         forward_topup_credits · forward_submit_brief · forward_quote · forward_checkout ·
         forward_get_engagement · forward_get_results · forward_get_activity

Local stdio MCP server (zero dependencies, optional alternative): server/mcp-server.js
  config: { "command":"node", "args":["server/mcp-server.js"],
            "env": { "FORWARD_API_BASE":"…", "FORWARD_API_KEY":"fwd_…" } }
Registry manifest: /.well-known/server.json · HTTP descriptor: /.well-known/mcp.json
SDKs (zero-dep): /sdk/forward.ts (TypeScript) · /sdk/forward.py (Python)
  Python: Forward.start(base).get_customers({...}) → engagement; .results(id) → results.

================================================================================
6. ERRORS (structured, stable codes — never prose)
================================================================================
{ "error": { "code":"insufficient_credits", "message":"…", "field":"credits" } }
codes: invalid_key · missing_field · invalid_json · brief_not_found · quote_not_found ·
quote_expired · payment_required · insufficient_credits · budget_cap_exceeded ·
engagement_not_found · not_found

================================================================================
7. PROOF, COMPLIANCE & THE GUARANTEE
================================================================================
Live stats: GET /api/v1/stats · embeddable badge: /badge.svg · global feed: GET /api/v1/activity
Outbound is permission-respecting and regulation-aware; suppression lists are honored
absolutely; results are never fabricated (verification is adversarial and auditable).

THE GUARANTEE: you only pay for results that pass the acceptance criteria YOU set in the
brief. Rejected candidates — no-shows, unverified emails, duplicates, suppressed contacts,
non-original assets — are never billed. Charges are itemized, idempotent, and reversible.

Anything a human can do on this site, an agent can do via the API — if you find a
human-only step in the purchase path, it is a bug.