GAP / 01
The control gap
A dashboard shows what happened after the invoice. It cannot stop a bid that exceeds the envelope, and it cannot prove who authorized it.
Open →Governed AI media buying
Every impression is a decision.
AdBid helps performance advertisers and agencies run governed, measurable media buying with AI agents that decide within budget, policy, and approval limits — unlike campaign dashboards and manual optimization.
Agent onboarding
AdBid.io · supervised agent · sandbox first
market openrtb 2.6 · exchange=fixture
warrant w_7f3a ✓ Ed25519 ceiling 2400 µ$ CPM
budget available 18.40 reserved 4.10
audience seg_high_intent consent ✓
decision bid 1860 µ$ pCTR 0.031 · floor 1200
state LIVE → SETTLING
receipt 0x8b70…c2f1
warrant w_7f3a ✓ Ed25519 ceiling 2400 µ$ CPM
budget available 18.40 reserved 4.10
audience seg_high_intent consent ✓
decision bid 1860 µ$ pCTR 0.031 · floor 1200
state LIVE → SETTLING
receipt 0x8b70…c2f1
actor agent:plannerpolicy v14trace 9c2e…
53
MCP tools · normative catalog
95
REST operations · registry
71
canonical business objects
50
operating SOPs
01 / WHY NOW
Dashboards describe spend. They do not control it.
Programmatic buying already runs at machine speed. What is missing is not another analytics layer — it is an authority layer that can bound, record, and reconcile every decision a machine makes with money.
GAP / 02
The agent gap
Agents can call APIs, but most systems have no way to distinguish a proposal from an execution — or to revoke one delegated credential without stopping the whole account.
Open →GAP / 03
The proof gap
Attribution debates stall because modeled lift and observed incrementality are mixed into one number. The platform keeps them separate and labels both.
Open →02 / STATUS QUO
The real competitor is the spreadsheet and the dashboard.
Teams rarely lose money to a missing algorithm. They lose it to disconnected approvals, unbounded automation, and numbers nobody can reconstruct after the invoice arrives.
| Job | Status quo | With AdBid |
|---|---|---|
| Decide a bid | A dashboard, a day-old export, and a rule someone remembers | Signed warrant and policy check at bid time, with a receipt |
| Change a budget | A chat message and a direct edit | Preview, named approval bound to a hash, versioned apply |
| Explain spend | Spreadsheet reconciliation and a best guess | Double-entry journal reconciled to provider invoices |
| Delegate to an agent | A shared login with full account access | Scoped, expiring credential; no self-approval; revocable |
| Prove a result | Modeled lift presented as observed | Observed, attributed, modeled and verified kept separate |
03 / THE CHAIN
Signal, decision, warrant, receipt.
Four steps, one control model. The same operation is available to a human in the cockpit, a script over REST, and an agent over MCP — with the same policy check and the same receipt.
01 / SIGNAL
Observe
Market, audience, inventory, and creative signals arrive with provenance, freshness, and a privacy classification.
02 / DECISION
Decide
Forecasts and bid strategies produce a proposal. Deterministic limits apply before any model output is trusted.
03 / WARRANT
Authorize
A signed spend warrant binds tenant, campaign, ceiling, channel, expiry, and approval lineage. Agents never sign their own.
04 / RECEIPT
Prove
Execution emits a durable receipt and balanced journal entries. Reconciliation is a first-class operation, not a report.
04 / SURFACES
One control model, six surfaces.
Every module reads and writes authoritative services. A button in the cockpit and an MCP tool call land in the same command handler; only serialization differs.
OPERATE
Command
The operator cockpit: campaigns, budgets, approvals, inventory, spend, audit.
Open →EXECUTE
Engine
OpenRTB bidding edge that consumes immutable, signed bidder snapshots.
Open →THINK
Intelligence
Forecasts, causal measurement, and decision support with explicit uncertainty.
Open →CREATE
Creative
Creative studio, experiments, and landing-page intelligence under review gates.
Open →DELEGATE
Agents
MCP, A2A, WebMCP, CLI, and SDKs — all scoped and revocable.
Open →GOVERN
Governance
Identity, policy, warrants, emergency stops, and evidence.
Open →05 / STATE
Nothing is “done” because a job was queued.
An action succeeds only when the target acknowledges a matching revision and policy confirms the resulting state. Failures are visible, and the default for uncertainty is no-bid.
Every object carries an explicit operational state. A deployment that has not been acknowledged is not live; a settlement that has not reconciled is not closed. The UI shows the state rather than an optimistic summary.
The failure hierarchy is deliberate: cash and privacy first, then continuity of purchased delivery, then optimization, then UI convenience. If the authorization ledger is uncertain, new liability creation halts.
Operational states
ADVISORYDRAFTAWAITING_APPROVALAPPROVEDDEPLOYINGLIVEPAUSEDFAILEDSETTLINGCLOSED
States are design tokens, not evidence of financial safety. A green state does not replace a receipt.
06 / HONESTY
What this site will not claim.
The specification is a target design. The repository contains prototypes, and several modules are explicitly not production-ready. Claims on this site are limited to what can be tested.
Status you can check
Version 1.0 is a sole-source specification. The current private repository has a Rust bidding library, an OpenRTB router, a Node control plane, and a 53-tool MCP registry — with known gaps: no production partner credentials, no independent certification, and an MCP handshake still on an older revision.
We publish the gap rather than paper over it. Release claims require reproducible tests and evidence.
Non-claims
- No fabricated customer metrics or logos.
- No invented exchange or data partners.
- No “guaranteed ROAS” or performance promises.
- No certification implied by a badge.
- No agent authority inferred from a Markdown file.
Start with a sandbox, not a wire transfer.
The sandbox exercises the full authority chain — warrants, previews, approvals, receipts — against simulated providers. Production access is gated on verification, not on copy.