Your brand's voice, in every agent. Day one.
Agents are about to scale across marketing. They are going to write the support replies, the sales follow-ups, the campaign pages, the investor notes, the developer docs, the internal comms. They are going to do it at machine speed, across surfaces no brand team has visibility into, and they are going to do it whether anyone briefs them or not. The question is whether they sound like you when they do.
Right now, almost none of them do. Every agent reaches for context the brand never provided. It pulls from cached pages, stale guidelines, a competitor's blog, the mean of the internet. The output is plausible. It is also generic. Six months in, the company sounds like everyone else. The differentiation the brand spent twenty years building gets averaged out one agent call at a time.
Encoded Brands fixes this once. We maintain an open standard for what a complete brand encoding looks like (BCP, the Brand Context Protocol, public at brandcontextprotocol.dev). We extract your actual voice into that format from the work you've already done. We host it on your domain and on our runtime. And we make it available to every agent your team deploys. Customer support, sales, paid media, internal copilots, anywhere. One source of truth your agents read before they speak.
The product is four parts: the Encoder (extracts your voice into BCP-conformant files machines can obey, available self-serve from $99 and enterprise from $25K), the Registry (hosts that BCP tree for every agent to read, with authentication and version control), the Monitor (watches what your agents actually say and flags drift), and the Encoded Brain (the pattern intelligence that gets sharper with every brand we encode). The business runs on a self-serve funnel that compounds into recurring Registry revenue and high-margin enterprise engagements. The spec is free. The runtime is the business.
The agents are coming. They are about to write everything your team used to write. The question is whether they sound like you when they do.
Every company is in the middle of an AI buildout right now. Customer support agents. Sales copilots. AI-generated campaign pages, RFP drafts, investor updates, developer docs, internal memos, account research. The output is going up by an order of magnitude. The volume is going up by two. Every one of those outputs is the brand speaking. And most of them are speaking generically.
The reason is simple. A brand is a set of decisions a person makes when nobody is watching. Use this word, not that one. Lead with this proof point, not that promise. Speak this way to a customer in week one, that way to a customer in year three. Those decisions live in the heads of senior people who have been at the company long enough to absorb them. They have never been written down in a way a machine could obey.
For most of the last forty years, that was fine. Brand guidelines lived in PDFs. Tone of voice lived in 90-minute onboarding workshops. The actual carriers of the brand were people, and people could be briefed.
That model breaks the moment people stop being the carriers. Three things happened at once.
We call the result brand drift. Every output is a little off. Six months in, the company sounds like everyone else. The differentiation the brand spent twenty years building gets averaged out one agent call at a time. By the time the CMO notices, the drift has happened across every surface — paid, owned, earned, internal — and there is no clean way to roll it back. The brand has been quietly rewritten by tools nobody briefed.
None of this is hypothetical. It is happening across every category right now. The companies that get ahead of it will set the voice of their agents on purpose. The companies that don't will find out what their voice became when a customer reads it back to them in a complaint.
One source of truth your agents read before they speak. Across every surface a machine speaks for you, paid and not.
Encoded Brands holds the canonical voice of your brand in one place, hosted on our infrastructure, available to every agent your team authorizes. Customer support reads it. Sales agents read it. Campaign tools read it. Custom GPTs and Claude projects read it. Internal copilots read it. One registry, one voice, every agent.
/.well-known/brand.md and /.well-known/brand/*. Free to read, free to author, free to consume. We maintain the spec at brandcontextprotocol.dev. Anyone can publish a BCP. Most won't on their own. That's the business.brand.json, we publish one alongside our BCP tree. Where they want richer depth, our runtime serves it.We are a software company that ships through high-touch engagements for the first two years. The Encoder is a real product. The Registry is real infrastructure. But the methodology in v1 is delivered by senior strategists in the room. That is how we earn the premium price for enterprise engagements, and it is how we capture the patterns that make the next encoding faster. Over time the Encoder carries more of the load and the strategist becomes the editor instead of the author. The methodology compresses into software as the corpus deepens. Both serve the same goal: your brand executing correctly at agent speed.
Where the ad-tech protocols meet our work, they are the skin. Logos, colors, the four-line tone object, the public file every agent can read. We are the soul. The narrative, the modulation, the claims, the patterns the brand refuses to be. Skin is published. Soul is hosted, authenticated, and paid for.
Four products. One business. Each one feeds the next.
A proprietary strategic compiler. Takes messy inputs — brand decks, founder interviews, sales calls, ten-year-old guidelines — and extracts the brand's actual DNA into a structured, machine-readable corpus. This is where the consulting practice meets the software product. Half methodology, half compiler.
Hosted serving + validation + attestation for every brand we encode. Customer BCP files served from our infrastructure, signed by us, validated daily for leaked secrets or stale content. Verified Publisher attestation that any agent can back-trace. AdCP-compatible brand_agent for paid media. Direct MCP endpoint for everything else. One service, many brands, multi-tenant. No editing GUI — corrections happen through Encoder re-runs or surgical admin edits.
Drift detection. Watches the actual outputs of authorized agents — ad creative, support replies, sales drafts, anywhere our context is being consumed — and reports back where the live output diverges from the registry. Closes the loop. Becomes the upsell.
The corpus that makes the Encoder smarter every session. Three logically separate tiers, each with different rules: Public (our frameworks and methodology), Patterns (anonymized cross-client learning, the long-term moat), and Client Vaults (each brand's content, isolated, never read into another session). The bright line between Patterns and Vaults is the architecture that protects every customer's trust.
The product surfaces are linked. The Encoder feeds the Registry. The Registry feeds the agents. The agents feed the Monitor. The Monitor feeds the Brain. The Brain feeds the next Encoder session. We sell three of them as products. The fourth (the Brain) is the engine of all of them.
A full picture of how context moves from a founder's head to a live agent and back again.
How a brand becomes encoded, step by step. The first session takes weeks. The hundredth will take days.
We collect every artifact the brand has produced about itself: founder decks, board decks, brand guidelines, the last twelve months of sales calls, the website, the last three campaigns, the help center, support transcripts, the last earnings call, the last all-hands. The Encoder reads all of it. The point is not to summarize. The point is to surface where the brand sounds distinctively itself and where it sounds like everyone else.
A structured, one-question-at-a-time pushback session with the founder and core team. The Encoder rejects vague platitudes. "Innovation" is not a value. "Authentic" is not a voice descriptor. We push until we get operational specificity: "We sound impatient with received wisdom but never with the customer." "We make concessions to clarity but never to brevity." "We can use jargon with developers and never with buyers."
The Encoder compiles the intake and interview output into the registry's structured corpus. This includes the AdCP-conformant brand.json public layer (logos, colors, basic tone, the standard fields) and the deep layer behind authentication: audience-modulated voice rules, claims with verification triggers, anti-AI patterns, surface-specific guidance, representation framing for AI search.
Every output is scored against a 25-point integrity rubric. If the brand name were removed, could the document only belong to one company? Are the dos and donts deep enough to bind an agent's behavior? Is every claim paired with verifiable evidence? Below 15 of 25, we reject as template slop. The session restarts. This is the bar.
We host the registry. We publish the brand's brand.json at their domain (with a pointer to our brand_agent) or accept their existing one. We hand over MCP credentials for the agents the brand wants to authorize. The Monitor goes live in shadow mode and starts watching.
Quarterly review. The Monitor's drift reports drive the cadence. Where is the brand actually slipping? Which agents are pulling context inconsistently? Where is the corpus thin? The Encoder returns. The Brain absorbs the patterns. The next brand benefits.
A reasonable customer will ask, before signing anything: "You're using our work to train your system. What stops our content from showing up in another brand's session?" The answer matters. We built the architecture around it.
The Encoded Brain is three logically separate stores. They have different rules for who writes to them and who reads from them.
The bright line is between Tier 3 and Tier 2. It runs in one direction only. Tier 3 content does not flow into Tier 2 automatically. It flows through human review, with anonymization rules that have to actually pass, and only after a pattern has been observed often enough to no longer identify any single source. If a CMO says "no patterns, period" — fine. We carve out Tier 2 contribution in the contract. The encoding still happens. The Brain just doesn't learn from that engagement. That is the customer's right.
For the self-serve Encoder (the $99 and $499 tiers we will offer publicly), Tier 2 is not in the access path at all. Self-serve customers get Tier 1 (methodology, rubric, frameworks) and their own Tier 3 (their content, their vault). The pattern intelligence is reserved for enterprise engagements where the contract, the legal coverage, and the strategist relationship are all real. This is also how we make the enterprise price defensible: the methodology is in the box at $499. The accumulated cross-brand intelligence is what you buy at $25K and up.
A public BCP file alone is a low-trust artifact. Anyone who controls a domain can publish anything they want at /.well-known/brand.md. DNS gets hacked. IT teams have keys marketing doesn't know about. Counterfeit brands stand up lookalike domains. Without a trust model, the spec is just convention with no defense against a bad actor publishing a fake BCP that defames a brand or rewrites their claims.
Encoded Brands operates the trust layer that makes BCP defensible at scale.
None of this is in the BCP spec itself today. The v0.2 spec update we are authoring adds the signing scheme, a registries field, validator requirements, and daughter file checksums. The spec defines the trust mechanics. We operate the runtime that delivers them.
The Encoder is not hard-coded to ask a fixed set of questions. It reads the Brand Context Protocol spec at the start of every session, sees what dimensions an encoding must cover (voice, values, boundaries, claims, representation, plus whatever the spec adds next), and shapes the interview around the live spec.
What this means in practice: when BCP v0.2 ships with a new daughter file — say, escalation.md for how the brand handles customer complaints — the Encoder automatically starts asking customers about escalation patterns at the next session. No code change. No redeploy. The spec is the contract; the Encoder is the dynamic consumer of it.
BCP defines what a complete encoding contains. The corpus defines how we ask. The spec is public and free at brandcontextprotocol.dev; the question library, the pushback prompts, the rubric, the pattern intelligence are all proprietary and live in our corpus. Anyone can theoretically author a BCP-conformant tree by hand. Most won't. The ones who try and decide they want help getting it right are our customer.
Pattern recognition in the Encoder is a staircase, not a single feature. Three kinds of patterns, each available at a different stage of the company.
Vocabulary patterns (day one). Words and phrases that are universally vague and never survive scoring. Innovation. Authentic. Premium. Trusted. Passion. Purpose-driven. Best-in-class. When a founder uses one of these in the interview, the Encoder pushes back: "What does that mean operationally? What would you sound like if you didn't use that word?" This is a vocabulary blocklist with structured follow-up. No data science required. It ships day one and makes every Encoder session feel sharp.
Failure shapes (v1, after ~20 engagements). Statistical patterns in how brands fail the 25-point rubric. "Brands that describe their voice as 'professional yet approachable' score below 12 of 25 on specificity." "Brands that list more than five values average 8 of 25 on enforceability." "B2B brands that fail to define a voice difference between sales and onboarding generate three times the drift in the Monitor." These are real observations across encoded brands. They require corpus depth to surface. They build over time, and that timing is fine.
Strategic moves (v2+, the long game). Patterns in how successful brands solve specific problems. "Fintech brands facing regulator scrutiny tend to over-index on trust signals, which weakens distinctiveness; the brands that maintained voice did X." This is the genuinely valuable pattern recognition — strategic moves the Encoder can suggest. It requires real corpus depth (50+ engagements minimum) and is where the moat lives long-term. The Brain compounds here, and a solo strategist cannot match it after a year.
Tier 1 patterns make v0 of the Encoder feel smart immediately, because every founder uses at least three blocklist words in their first answer. Tier 2 patterns earn credibility once we have a portfolio. Tier 3 patterns are the reason the company is valuable in five years.
The Registry is only useful if the integration to your existing AI stack is easy. Most of it is.
Direct MCP integration. Claude, ChatGPT Enterprise, Cursor, and any AI tool with MCP support adds a server with a single URL. Customer goes to Settings → Add MCP server → paste their Registry URL → done. The agent now consults brand context on every relevant query. As easy as adding a Slack integration.
Custom GPTs and Claude Projects. The customer puts the registry URL in the system prompt. Three lines of text. Their internal team's prompt now includes "consult brand context at registry.encodedbrands.com/[brand-id] before generating." No engineering required. Documentation and copy-paste snippets carry the load.
Marketing tools with native brand context support. Canva, Adobe, and the broader ad-tech ecosystem are building agentic infrastructure that reads brand identity from public endpoints. Where those partners read a brand.json file, we publish one. Where they want depth, our Registry serves it through the same pointer. The integration is automatic for anything reading the ad-tech protocols. This is the only place Skin and Soul meet in practice.
AI customer support and sales platforms. Intercom Fin, Decagon, Sierra, Salesforce Einstein. These do not have MCP support yet, so the integration is the customer's CX or RevOps team adding the registry content to the agent's system prompt or knowledge base. Two-to-four-hour engineering task on the customer side. We make it easier with documentation, snippets, and a "configure your support agent" guide.
Custom enterprise agents. Internal tools built by the customer's engineering team. They write the integration themselves. We provide developer docs and a reference implementation. This is the longest tail and the hardest sell, but it is also where the largest enterprise customers actually live.
The promise to the customer: configure once. Every agent stays on-brand forever. The work behind it is an integration playbook we ship with the registry, plus the account manager support that comes with enterprise tier. At self-serve tier, the customer gets the playbook and a copy-paste guide for the top five surfaces. At enterprise, we configure their actual agent stack with them.
The defensibility is not the file. It's not the methodology. It's the corpus that gets sharper every time we use it.
Each Encoder session does three things. It compiles a brand. It scores itself against the rubric. It writes back into the Encoded Brain whatever pattern produced the score. After a hundred sessions, the Encoder knows how a fintech brand fails differently from a CPG brand, how a B2B SaaS brand modulates voice between sales and onboarding, which X-not-Y rules produce drift in support and which ones hold up.
This is the part of the business that does not exist on day one. It accrues with each session. By session 50, the Encoder pushes back on patterns we have seen fail before. By session 200, no human strategist alive has the same pattern library. That is the long-term moat. The Registry is the recurring revenue. The Brain is the company.
Five revenue lines, arranged so each step is roughly 5-10x the previous one. The funnel actually compounds when the step gaps stay reasonable.
brand_agent MCP server on our infrastructure. Multi-tenant isolation. Capped call volume (5K reads/mo at entry, scales up). Standard schema. Self-service onboarding, community support, no SLAs. The on-ramp from one-time Encoder purchase to recurring revenue. Many customers will live here forever and that is fine — they are cheap to serve and they are public brand_agents on real domains.The steps are roughly: $99-$499 one-time → $49-$199/month → $25K-$100K one-time + $2K-$5K/month. Each jump is 5-10x, not 50x. A self-serve customer that buys the Encoder and stays on the standard Registry for two years has paid us $1,500 to $5,000 cumulative. That is real money at volume, and it is also the customer who eventually upgrades to enterprise when their team grows or their first AI customer support agent goes off-brand in a memorable way.
The tier separation also explains the price gaps. At $499 you get the methodology applied to your brand. At $25K and up you get the methodology plus the accumulated pattern intelligence of every previous enterprise engagement, plus a strategist accountable for the result. Same product surface, different intelligence behind it. The methodology is in the box. The Brain is what you pay enterprise for.
The Registry is high-margin at scale and infrastructure-heavy to get there. A hosted multi-tenant MCP server with authentication, Tier 3 namespace isolation, audit logs, and uptime guarantees is real software that needs a real engineer to build and a real bill to pay. Realistic ramp: one contract engineer in month one to ship v0 (brand_agent + auth + Tier 3 isolation), a second engineer when we cross 20 paying Registry customers, the Monitor and audit logging as v1 work after Cannes. At $2K-$5K per enterprise Registry customer, 10-15 of them covers a full-time engineer plus infrastructure. The math works at scale. The runway to get there is a real number we have to fund.
We are not betting on becoming a public registry. We are not betting on running a foundation. We are not betting on Zefr or any single ad-tech partner. We are betting on being the deep brand context operator that fifty brands have hired by Cannes 2027 and three hundred have hired by Cannes 2028. Operators win this category. Standards bodies get acquired by operators.
Five weeks to Cannes. Sequenced honestly. Built around what we can actually defend in a press conversation and what we can actually ship with the team we have.
brandcontextprotocol.dev. Penny owns the fixes: (a) verify the registry at registry.brandcontextprotocol.dev actually resolves, or ship a one-page version listing encodedbrands.com as the inaugural entry; (b) fix the broken authoring/llm-prompt.md reference; (c) refresh site copy so BCP is positioned as the public file format Encoded Brands maintains, not the company's main product; (d) soften "maintained by Encoded Brands and the community" to drop the "and the community" claim until there is one. Half day of work. Has to happen before any press-facing conversation.brand_agent + auth + Tier 3 namespace isolation. Monitor, audit logs, and custom schema extensions are v1 work after Cannes. Named individual signed by end of week. Without this hire, "stand up the Registry" is fiction.brand_agent implementation spec end-to-end. Done as of this memo. The MCP tools we need to implement are get_adcp_capabilities, get_brand_identity, and (optionally) get_rights. The depth fields beyond AdCP's schema are ours to define.registries field in brand.md frontmatter, daughter file checksums for tamper detection, and validator requirements (what gets scanned, what gets rejected). Penny pushes the spec update to brandcontextprotocol.dev. Trust model section credits Chris as advisor.brand_agent MCP server. Multi-tenant. Authentication and Tier 3 isolation on day one. Two pilot brands deployed by end of month.brand_agents deployed on real domains. Pilot customers from Nick's and James's networks. Even three deployments puts us as the largest depth-layer deployer in the world — AdCP's registry has zero real deployments today.brand_agents already deployed on real domains. Self-serve product in beta. Talking with the AdCP working group about depth extensions for late 2026."