Consul: The Dream

A vision document for the best possible product, written without regard for how hard it is to build.

The one sentence

Consul is a single named entity that runs your professional life in the background, learns you over months, takes consequential action with calibrated trust, and is wrong about you in ways you can fix yourself.

Every word in that sentence is doing work. The rest of this document unpacks each claim and says what it means for the product, what it rules in, and what it rules out.

Claim 1: A single named entity

Consul is not a tool. It is not a feature. It is not a workflow platform. It is a persistent character with a name, a voice, a memory, and a relationship arc.

The user hires Consul the way they would hire a person. There is an onboarding week where it is learning. There is a three-month mark where it starts anticipating. There is a one-year mark where it knows things about the user the user has forgotten. The product is the relationship, not the feature set.

This is the thing OpenClaw understood that the rest of the market does not: people do not want a Swiss Army knife. They want someone. The naming, the voice, the consistency across surfaces, the way Consul refers to past conversations — these are not decoration. They are the product.

The implication: there is exactly one Consul per user. Not a team of bots. Not a fleet of specialists. Not a marketplace of skills the user assembles. One assistant, one identity, one memory, one voice. Specialization happens inside Consul as private capabilities the user never directly manages — the same way a human EA "specializes" by being good at travel without you hiring a separate travel agent.

The "team of named specialists" framing is a developer's mental model of an org chart leaking into a consumer product. Real principals do not want to manage a team. They want one chief of staff who manages the team for them.

Claim 2: Runs your professional life in the background

Background is the operative word. The best version of Consul is the one the user forgets is running for hours at a time, and is then pleasantly surprised by.

That requires four things no current product has:

A real heartbeat. A loop on a tight cadence — fifteen minutes, not thirty — running against a watchlist the user wrote in plain English. Not scheduled jobs. Not cron. A continuously-running agent that wakes, looks, decides, and either acts or goes quiet again.

Multi-source situational awareness. Inbox, calendar, Slack, Drive, location, time of day, the user's last activity. Consul knows you just landed because your phone changed time zones and your calendar said you would be in the air, and it preemptively drafts the "I'm in town" email to the person you are meeting tomorrow. The model already exists. Nobody has built the state layer that feeds it correctly.

Silence by default. The vast majority of heartbeat ticks produce zero outbound communication. The user notices Consul because of the quality of the few interruptions, not the volume. A Consul that sends three messages a day is a Consul that gets muted on day four. A Consul that sends one perfectly-timed message every other day is one the user cannot live without.

Cross-context stitching. The email from Tuesday, the calendar invite from Wednesday, the Slack thread from this morning, and the document you opened yesterday are one situation in Consul's head, not four. Today's products treat each surface as its own island. The dream version treats them as fragments of a single story it is responsible for understanding.

Lindy and Righthand and Base44 are all reactive-with-cron dressed up as proactivity. The dream is continuously situationally aware, which is a different category of system, not a more-frequent version of the same one.

Claim 3: Learns you over months

This is the part everyone fakes and nobody builds. Real personalization is a compounding loop, and a compounding loop has three pieces current systems do not have:

An editable model of you. A prose document, viewable and writable by the user, that contains "what Consul believes about me." Not chat history. Not vector embeddings. A living file:

Hates Tuesday morning meetings. Calls his wife Em, never Emily. Inbox-zeros on Sundays. Considers anything from marcus@acme high-priority. Is trying to delegate more — push back gently when he tries to do something he could hand off.

The user can read it. Correct it. Delete things. Add things. Wrongness becomes a fixable bug instead of a permanent annoyance. This is the single most important architectural commitment in the entire product, and it is the one every competitor has explicitly decided against because opaque memory is easier to ship.

Preference learning from the deltas, not the ratings. When the user edits a draft Consul wrote, Consul learns from what changed. When the user ignores three suggested follow-ups in a row, Consul learns to suggest fewer. When the user always reschedules things Consul booked on Monday afternoons, Consul learns Monday afternoons are sacred. This is a real per-user learning loop, not a prompt that says "remember the user's preferences."

Episodic memory with retrieval, not summarization. The user should be able to ask "what did I tell you about the Henderson deal three months ago" and get the actual answer — the actual sentences from the actual conversation, not a vibes-based reconstruction. Structured event logs, real retrieval, no lossy compaction.

The compounding is the moat. After six months, Consul is irreplaceable in a way ChatGPT, Lindy, Righthand, and Base44 structurally cannot be. Not because the model is better. Because the model of you is better.

Claim 4: Takes consequential action with calibrated trust

Most current products are stuck in "draft for approval" because they are afraid of being wrong. The dream version is genuinely brave: it has graduated autonomy that earns more rope by demonstrating competence.

• Week 1. Drafts everything. Sends nothing.
• Week 4. Sends low-stakes things automatically — acknowledgments, calendar confirms, "got it, will revert." Drafts everything else.
• Month 3. Handles entire categories autonomously — scheduling with known contacts, expense filing, routine follow-ups. Still drafts anything novel or sensitive.
• Month 6. Takes initiative. "I noticed you and Marcus have not talked in six weeks and you said you wanted to keep that warm — want me to suggest a coffee?"
• Month 12. Acts as a peer. "I pushed back on the Henderson scope creep on your behalf. Here is what I said and why. Here is the reply. You can override."

The trust escalation is itself the product. Not a setting the user toggles. A relationship that develops. The user feels Consul earning more responsibility the way a manager feels a new hire earning more responsibility — through demonstrated judgment on small things first.

Three components of the trust loop most products skip:

Plans before actions. For any multi-step work, Consul shows the entire intended sequence as one reviewable artifact, not as fourteen individual approvals. The user reads the plan, edits the plan, approves the plan once, and the plan executes.

Receipts after actions. Honest about uncertainty. "I sent this. I was 80% sure you would want this tone. If I am wrong, tell me and I will learn it."

Reversibility as a first-class concept. Every action Consul takes has an undo path Consul knows about and will execute on request. Unsend that works for a window. Move that meeting back works always. Pretend you never saw that email updates Consul's model so it does not act on it again.

Claim 5: Wrong about you in ways you can fix yourself

This is the philosophical claim that separates Consul from every other product in the space.

The best AI assistant is not the one that is right most often. It is the one whose wrongness is legible and correctable.

Lindy is wrong in ways you cannot see — a draft sounds slightly off and you do not know why. ChatGPT is wrong in ways you cannot fix — it forgets the correction next session. Base44 is wrong in ways you cannot audit — a workflow misfires and you do not know which trigger caused it. Consul is wrong in ways the user can inspect, understand, and repair through normal conversation.

The model of you is a file you can open. Decisions are explainable in plain English. Mistakes produce visible changes in Consul's user file that the user can see happened. "You drafted that too formally" is not a complaint that disappears into the void — it is a sentence that updates a line in a document the user can read.

This is the property that makes long-term use survivable. Every other AI product accumulates frustration over months as small wrongnesses compound. Consul accumulates fit over months because each correction sticks and is visible.

What Consul is not

The dream is defined as much by its refusals as by its commitments.

Not a platform. No marketplace. No third-party skills. No plugin ecosystem. No developer SDK. Every capability is built and curated by the Consul team. This is the Apple move, not the Android move, and it is right for this category because trust cannot be marketplace-distributed.

Not horizontal. Does not try to be a sales agent or a marketing agent or a support agent. Only does the work of an executive assistant for an individual professional. The horizontal version is Base44, and it will be shallow forever.

Not a chatbot UI. The chat surface exists because that is how humans talk to assistants, but the product is not "a place you go." It is a presence that reaches you on iMessage when needed, sends a morning brief to your inbox, and otherwise stays out of your way. The web app is for inspection and configuration, not for daily use.

Not multi-user inside one account. One user, one Consul. No shared workspaces. No team Consul. The relationship model breaks the moment you try to make one assistant serve two principals. That is a different product — Chief of Staff for a team — and it should be a different product.

Not free. The dream version costs $200-500 per month, because that is the only price point at which the unit economics support real always-on infrastructure, real trust-stack engineering, and real human-quality model usage. Free or$20/month forces architectural compromises that kill the vision. Pricing is a product decision. Cheap kills this product.

The one-paragraph version

The best possible Consul is a single named entity the user hires for $300 a month, that runs continuously in the background watching their professional life across email, calendar, and messaging, that maintains an editable plain-English model of who the user is and what they care about, that takes graduated autonomous action calibrated to demonstrated trust, that shows its plans before acting and its receipts after, that gets visibly better at the user over months in ways the user can audit and correct, that lives on iMessage and email and stays silent unless it has something worth saying, and that — six months in — the user cannot imagine working without and cannot describe to a friend without sounding like they're describing a person.

Who this is for

The dream version is not for everyone, and pretending otherwise weakens it.

It is for the professional who currently pays — or wishes they could pay — $8,000 a month for a human executive assistant, and would happily pay$500 a month for 80% of that value. It is for the person who wants a relationship with their software, who values long-term fit over immediate utility, who is willing to invest a week of corrections to get a year of compounding rightness.

It is not for the person who wants Lindy-but-cheaper. That market is larger and will be served by a race-to-the-bottom that Consul should not enter.

The bet is that the smaller market is defensible and the larger market is not. The bet is that one entity that knows you is a category that does not yet exist, and the company that builds it first owns the relationship-with-software space the way the iPhone owned the relationship-with-phone space — by committing to a vision the rest of the market thought was overbuilt until the moment it was obviously right.

The primitive spectrum

The dream is abstract until you map it against the actual building blocks that make it possible. Below is every primitive we have identified across OpenClaw, NemoClaw, and the broader agent-infrastructure space, scored against the real systems in the market — including Consul as it exists today and Consul as the dream describes it.

Legend: ✅ strong · ⚠️ weak or partial · ❌ missing · ➖ not applicable

Runtime primitives — the shape of the system

Ambient primitives — the thing that makes it feel alive

Memory primitives — the compounding moat

Trust primitives — the brave part

Safety primitives — the invisible moat

What the table says

Count the columns. Chat products implement roughly three of twenty-eight primitives meaningfully. Lindy implements about four. Righthand implements about five with one clear standout. Base44 implements about six. OpenClaw implements about twenty. NemoClaw dominates the six safety primitives and does not compete on the rest. Consul-today implements about six, and almost all of them in weak form.

The dream version is the only column with ✅ across all twenty-eight rows, and that is not an accident — it is the definition. Every claim in the one-sentence vision maps to a specific cluster of primitives above. Runs your professional life in the background is the Runtime and Ambient clusters. Learns you over months is the Memory cluster. Takes consequential action with calibrated trust is the Trust cluster. Wrong in ways you can fix yourself is primitives 12, 18, 19, and 20 specifically. A single named entity is primitive 15 holding the whole thing together.

Three patterns worth naming out loud:

The ambient cluster is where competitors are weakest. Primitives 7 through 11 — heartbeat, silent-success, quiet hours, standing objectives, cross-context stitching — are nearly empty across the entire market. This is the cluster that separates "assistant you text" from "chief of staff who notices things," and nobody is building it. This is where Consul wins on user feel.

The safety cluster is where competitors are most dishonest. Primitives 23 through 28 are either missing entirely or handwaved with words like "sandboxed" and "secure by default" that do not survive a technical audit. NemoClaw is the only column with real marks here, and NemoClaw is not a consumer product. This is where Consul wins on enterprise sales, compliance review, and the Wilkinson-quote gap.

The memory cluster is the compounding moat. Primitives 12 through 16 are the ones that get stronger over time rather than at launch. A competitor who ships without these can never catch up to a competitor who ships with them, because the advantage compounds with every week of user interaction. This is where Consul wins on retention — not on day one, but on month six, when every other product still feels like day one.

The twenty-two primitives Consul today is missing are not evenly weighted. The ambient cluster is existential to the category claim. The memory cluster is existential to the moat. The safety cluster is existential to the trust positioning. The trust cluster is existential to the "brave" part of the product. Missing one cluster means shipping a different product. Missing all four means shipping Lindy.

A closing note about the gap

Two honest things to say out loud, because the dream is only useful if it is named alongside the distance from here to it.

First. This product is not a stretch version of Consul-today. It is a different product that shares a name. Consul-today is a polite intern with strong human-in-the-loop. The dream Consul is a persistent character with an editable self-model and graduated autonomy. The gap is not features — it is the runtime, the memory architecture, the trust loop, and the philosophical commitment to one entity per user. Getting from here to there is not "ship more skills." It is a rebuild of the spine.

Second. Every primitive in this document is buildable. None of it requires a model breakthrough. The heartbeat is engineering. The editable memory is engineering. The graduated autonomy is engineering and design. The trust receipts are engineering and copywriting. The credential broker is engineering and security work. The dream is not blocked on AI progress. It is blocked on whether the team is willing to build the unglamorous infrastructure that makes the magic possible.

That is the whole bet. The model is a commodity. The runtime is the moat. The relationship is the product.

Build the runtime. Earn the relationship. Be the product nobody else is willing to build.