One shared record that people and agents both work from.
Hub is an agent-first operations platform: projects, tasks, clients, signals, evidence, knowledge and agent runs in one record, behind one permission model. People work in it without a model. Agents work in it through a keyed, scoped, revocable API.
The record is the interface. A chat thread is not.
A durable shared record sits at the centre. Jobs (bounded, single-output steps run by agents or people) read from and write to it. Dependencies are drawn only where a real handoff exists. And a fixed, fail-closed chain of checks, not a permissions checkbox, sits between any agent's proposal and any action touching money, a live system, or a person outside the organisation. It is not an autopilot, and doesn't claim to be.
You stay in charge
Any action that spends money, publishes, or writes to an outside system needs a named human's yes on one exact payload. Edit the payload and the approval dies with it.
You can see what it did
Append-only audit rows for every attempt, allowed and refused. A logged refusal names the strongest reason.
One shared record
The record is the interface, not a chat thread or any single tool's private database. Eight tools become eight views of one thing.
The cost is visible
Tokens and money on the job row as it runs, per model. A cost that isn't known is left blank, never shown as zero.
Nine words, used the same way everywhere.
Every diagram, every table and every row on the board uses these, and the right-hand column matters as much as the left.
- Task
- A human-visible commitment or piece of work.Not every internal model call.
- Job
- One bounded step, one owner, one typed output.Not a permanent persona.
- Arrow
- A real dependency: the next job needs the last one's output.Not just chronological order.
- Run
- One execution of a graph, for one task.Not the reusable workflow definition.
- Gate
- A named human decision on an exact version, required before continuing.Not a status label nothing enforces.
- Fixed action
- Executed personally by a human, because it spends money or can't be undone.Not an agent action with a rubber stamp.
- Library
- Written by a person (brand voice, do and don't-say). Read-only to runs.Not agent-generated content.
- State
- Data a system produced or holds: the record, a source feed.Not a model's private context window.
- Artefact
- A run's output for a person to check.Not live before someone approves it.
What's in the box, with the board's own count beside each one.
Capabilities are declared per area and each carries the file or row that proves its state. Built, partial and missing are the board's words, not ours. Updated 1 September 2026.
01 · 54 of 61 board rows done
Projects & tasks
Boards, the task table and drawer, subtasks, comments, estimates and actuals, time tracking, My Tasks. The most finished area.
- builtProjects board
- builtTask table and detail drawer
- builtTask comments and activity feed
- builtSubtasks and the task tree
- builtEstimates and actuals
- builtTime tracking
- builtWork logextra
- builtThe corrected task model — nine categories, five statuses, four priorities, stage
- builtMy Tasks — the personal task surface
- builtSections — the board’s workflow lanes
- partialProjects table — rank, journey stage and spend columnsextra
- partialAutomations write the corrected task model
- missingThe reset — one clean task corpus
- missingTasks linked to agent runsextra
02 · 30 of 42 board rows done
Clients & CRM
Directory, client record, contacts, leads pipeline, interactions ledger, tickets and enquiries, and the journey engine that writes follow-up tasks.
- builtClient directory and board
- builtClient record
- builtPeople-for-review queue
- builtInteractions ledger and manual log
- builtClient invites and the login roster
- builtClient attention and nags, with resolve and snoozeextra
- builtContacts and people
- builtTickets and enquiries
- builtLeads pipeline
- partialIndustry and segmentationextra
- partialJourney engine — welcome sequences, follow-ups, stall engine
03 · 17 of 22 board rows done
Client portal
What the client sees: track record, reviews and approvals, projects, reports, docs and account. Real records through row-level security, open since 1 September.
- builtPortal shell and navigation
- builtTrack record
- builtReviews and approvals
- builtPortal projects — all projects, growth roadmap, work log
- builtPortal account — ad hoc hours, business details, plan, connections, settings, invoices
- builtDocs paneextra
- builtSeated-client dock drawerextra
- partialChannel Workbench
- builtOverview
- partialWeekly and monthly reports
- partialClient targets display
- builtPrinted / PDF outputextra
- partialClient login to the portal is open
- partialBook a meetingextra
- ruled outWorkbench future channels — clarity, ads, meta, testing, local, email
04 · 16 of 20 board rows done
Knowledge & skills
The block-editor wiki, snippets, the client library (brand, voice, design system), and the skill contract that agents work from.
- builtDocs / wiki with block editor
- builtSnippets library
- builtClient library — brand, voice, design system, Drive
- builtCo-editingextra
- partialSkill contract and release control
- missingSkill body readable by a running job
- missingSWOT catalogueextra
- partialSkills on the task page — what a run may useextra
- ruled outSkills catalogue surface
- ruled outSkill studio
05 · 36 of 66 board rows done
Agents & workflows
The approvals desk, the model resolver, the worker fleet, the run ledger and the gates. The area the whole platform turns on.
- builtApprovals desk
- builtModel picker, resolver and routing policy
- builtCommand-bar ask assistantextra
- builtOperator brief handed to a worker
- partialRunner spine and fairness
- builtWorker fleet — dispatcher, executor, lease reclaim
- partialTask page — where a person meets a run
- partialBudget caps enforced
- partialApproval escalation reaches a human
- partialRun graphextra
- partialConnections context pack — what a routed agent is toldextra
- partialRun recap — the memory loopextra
- missingRun ledger — anything written at all
- missingA job running start to finish
- missingFleet run-health roll-upextra
- ruled outScheduled sessions
- partialAutomations registry, library and silence alarmextra
06 · 8 of 21 board rows done
Signals & evidence
Site health, refusal-safe client numbers, the notifications hub, the ingest workers, and a freshness stamp on every displayed figure.
- builtSite health
- builtRefusal-safe client numbers
- builtNotifications hub on live data
- builtPortfolio boardextra
- builtExecutive rollupextra
- partialData ingest — the workers that produce every number
- partialFreshness stamp on every displayed number
- partialWebsite performance panelextra
- partialMonthly reports snapshot their numbers when the month closesextra
- ruled outTransforms beyond the first board
- ruled outWarehouse layer
07 · 27 of 34 board rows done
Security & access
Row-level security across every client-reachable table, the role model, step-up auth, secret custody, agent API scoping, the audit trail and the permission floor on approvals.
- builtRLS across client-reachable tables
- builtRole model and team gating
- builtStep-up authentication for admin writes
- builtGuest and share-link access control
- builtSecret custody
- builtAgent API scopingextra
- builtWorker secret and tool isolationextra
- builtAudit trail — who did what, reviewable
- partialUploaded-file and storage access control
- partialWorker spawn fenceextra
- partialGrant posture — deny by default, no latent anon reachextra
- builtLegacy workers off the master key
- partialWorker confinement on the ops machineextra
- builtPermission floor on approvals
08 · 143 of 170 board rows done
Platform & estate
The test battery, the parity harness, CI, replay tooling, the deploy paths, backups and alerting. Reported separately because no Hub user touches most of it.
- builtTest battery and gatesextra
- builtDesign parity harnessextra
- builtRoadmap board and agent interfaceextra
- builtContinuous integrationextra
- builtDatabase replay and rollout toolingextra
- builtFront-end rollbackextra
- builtProduction deploy path
- partialThe ops-machine worker deploy path
- partialObservability and alerting
- partialOperational runbooksextra
- partialCredential recovery if the ops machine diedextra
- partialBackup and tested restore
Three relationships, and none of them is integrates-with.
Replaces: where the Hub record becomes the system of record
| Category | What it holds | Tools it stands in for |
|---|---|---|
| Projects & tasks | What is happening, who owns it, what is next | Asana, ClickUp, Monday, Trello, Jira, Linear, Basecamp |
| Sites & pages | Build, publish and edit the client's site | WordPress, Squarespace, Wix, Webflow |
| Review & mark-up | Comment on a page or a video, approve a version | Frame.io, Markup.io, Filestage, Pastel, BugHerd |
| Client portal | Reports, approvals, documents, progress | SuiteDash, Copilot, Moxo, HoneyBook, Dubsado |
| Knowledge & brand | Methods, brand, assets, operating instructions | Notion, Confluence, Frontify, Canva |
Connects to: reads, drafts, and the human gate on every write
| Platform | Reads | Writes | Human gate on the write |
|---|---|---|---|
| Google Ads | Spend, performance, search terms, quality signals | Campaigns, ad groups, copy, budgets, created paused | Enabling spend · any budget change |
| Meta Ads | Performance, audiences, creative results | Draft campaigns and creative | Publishing · enabling spend |
| GA4, Search Console | Sessions, conversions, queries, coverage, CTR | Nothing | Read-only by design |
| Tag Manager | Container config, tags, triggers, variables | Staged container versions | Publishing the version |
| Clarity, Semrush, Ahrefs | Behaviour, rankings, keywords, backlinks | Nothing | Read-only by design |
| HubSpot, Salesforce, Pipedrive, Zoho | Contacts, deals, pipeline, activity history | Contacts, notes, stage changes | Anything outbound to a person |
| Mailchimp, ActiveCampaign, Campaign Monitor, Klaviyo | Lists, segments, campaign results, engagement | Draft campaigns, segments, schedules | Sending · always human |
| Google Drive, Dropbox | Folder structure, documents, assets | Deliverables, reports, exports | Writing into client-shared folders |
| Gmail | Classified threads, client history | Drafts only | Sending · always human |
| Xero | Invoices, payments, contacts | Draft invoices | Approving and sending |
| GitHub, Vercel | Repository, build and deploy status | Pull requests | Merge and deploy |
| Google Calendar | Availability, bookings, client events | Holds and draft invites | Sending an invite |
Runs on: the engine, not something replaced
| Group | What it is | Tools |
|---|---|---|
| Frontier models | The heavy reasoning: planning, writing, judgement | Claude, ChatGPT; Gemini, Kimi and Qwen planned |
| Local model options | The free floor, for triage and high-volume simple work | Ollama, Gemma, Qwen (planned after local proof) |
| Data and runners | Moving work between systems, and holding what comes back | Airbyte, Nango, n8n, Supabase |
Most of the value lives in the first tier. None of the blast radius does.
Transitory
Runs freely. Chains freely. No gate.
Drafts and research, inert until a human uses them: SEO digests, copy drafts, page mock-ups.
Reversible
Gate: human approval. The pull request is the gate.
Publish a page, commit to the repo, update site copy. Always staged, always approved, always undoable.
Fixed
Recommendation only. A human executes.
Money and external state: changing ad spend, anything touching a client account with no undo.
Five things we refuse to build first.
- An unrestricted canvasRead a run before you can author one. A bounded grid, not a whiteboard.
- A second task storeOne record. The moment there are two, one of them is wrong.
- An opaque autopilotAutonomy grows in breadth, not in risk. Low-risk action classes earn lighter review on evidence.
- One agent for every stepA hundred agents is queue depth and a capability registry, not a hundred open sessions.
- Broad write accessMinimum practical scope, client isolation, revocability, and an attributable record. Every connection.