Intelligence
Not a chat box with a model behind it. It is the Claude Code CLI running as an agent with tools, so it can read your own API, run a command and come back with the answer instead of a guess.
Agentic marketing operations
One install. The whole operation runs inside it.
MEIQR is the platform behind an agency that does not scale by hiring. A fleet of agents reads your accounts when you run it, the analytics stop being assembled by hand, and every write action waits for a human. Installed and configured for one owner, not sold as a seat in someone else tenant.
- 161
- API routes
- 10
- agents in the fleet
- 8
- specialist squads
- 13
- analytics surfaces
The shape of it
Multi-tenant, inside your own install
Every organisation and every project gets its own address, with row level security in the database keeping one client out of another. The capability is the same one an agency needs to serve twenty accounts, except the install belongs to you and the data never leaves it.
There is no shared platform to log into and no subscription seat to lose. The product is the installation.
routing
# one install, N clients
/o/:org/p/:project
# the surfaces under a project
/o/acme/p/launch/intelligence
/o/acme/p/launch/mission
/o/acme/p/launch/analytics/meta
/o/acme/p/launch/funnels
# row level security, per tenant
RLS enabled on every table
Modules
Nine surfaces, one operating picture
These are the items in the sidebar of a running install, not a roadmap. Each one is a screen a human opens and a set of routes an agent can call.
Second Brain
The memory graph. Notes, decisions and project history become a searchable graph the agents read before they act, so context stops living in one person head.
Mission Control
Every agent run in one place, with what it proposed and what it spent. Write actions arrive as pending approvals, never as a done deal.
Command Center
The daily surface of one project: what moved, what is stuck, what needs a human today.
Tasks
Work items the agents pick up and close, with the run log attached to each one.
Funnels
Funnel builder with products and checkout config, plus lead capture, conversion metrics and an event timeline per lead. The light CRM for the client who has none.
Agents
The roster itself. Each agent has a role, a cadence, a permission tier and an approval rule you can read before you turn it on.
Squads
Eight specialist squads orchestrated by a strategy lead on Opus, for work that needs more than one head: research, traffic, dev, copy, content, audiovisual, infra.
Skills
File based skills discovered straight from the .claude directory. Adding a capability is adding a folder, not shipping a release.
Analytics
Thirteen surfaces, one source each
Paid, organic, attribution and pipeline in the same place, each reading the client own account. A dashboard with no credential behind it does not guess: the API answers with an error and the screen says what is missing.
That rule exists because the opposite is what quietly ruins a report. A number that looks plausible and belongs to someone else is worse than an empty chart.
- Overview the whole account on one screen
- GA4 Traffic sessions, sources, conversions
- Meta Ads campaigns, ad sets, ads, spend
- Google Ads performance and cost
- SEO organic position, CTR, ranking pages
- Keyword Scout where the demand actually is
- Market Demand category level interest
- Demand Forecast what next month looks like
- Projections budget against outcome
- Attribution which channel earned the credit
- Opportunities gaps worth a decision
- Users who is in the account
- CRM and Enquiries pipeline and source of each enquiry
The fleet
Ten agents with a role, a cadence and a permission tier
Read tier agents run on their own and report. Action tier agents can change something, and those arrive as an approval in Mission Control before anything happens.
| Agent | Role | Area | Cadence | Tier |
|---|---|---|---|---|
| Gus | Chief of staff | Ops | 07:45 daily | read |
| Ada | Research and intel | Intelligence | 09:00 daily | read |
| Sol | SEO and AEO | Growth | 08:00 daily | read |
| Zip | Social monitor | Intelligence | Fri and 08:00 daily | read |
| Mia | Meetings | Ops | 20:00 and 21:30 | read |
| Vic | Vault keeper | System | 18:50 and 08:30 | action |
| Leo | Content lead | Content | 06:12 daily | action |
| Edu | Editor | Content | on assignment | action |
| Rex | Paid ads | Growth | on assignment | action |
| Kai | Agency delivery | Clients | on assignment | action |
And eight squads for work that needs more than one head
Strategy
Orchestration, briefing analysis, dependencies
Research
Deep web research, competitors, audience
Traffic
Meta Ads campaigns, A/B testing, scaling
Dev
Landing pages, integrations, CRO, deploy
Copy
Page copy, email sequences, cold outreach
Content
Scripts, social posts, editorial strategy
Audiovisual
Motion graphics, AI video and audio, VFX
Infra
VPS, Docker, Supabase, CI/CD, DNS and SSL
meta ads tool grants
# what the agent is allowed to hold
TIER1_READ_ONLY = 18 tools
TIER2_ACTIONS = 13 tools
create_campaign, create_ad_set, create_ad_creative
BLOCKED_TOOLS = every delete and every token operation
# the rule in one line
creating is allowed, deleting never is
Paid media
A campaign built by conversation, paused by design
Ask for a campaign and the agent assembles it in the client own ad account: campaign, ad set, creative. It stops there, paused, waiting for a person to press go. The permission list is explicit and deletion is not on it.
A system that can only err on the side of doing less is a system you can leave running.
Second brain
The memory the agents read before they act
Notes, decisions, playbooks and project history become a graph with search and an in app reader. It is the difference between an agent that asks you the same question every week and one that already knows what was decided in March, and why.
The graph moving behind this page is that idea, running. Nodes are memories and the edges are what connects them.
Architecture
Three moving parts, all of them yours
Front
A static dashboard on the edge. It is the part the client opens, and it costs nothing to keep online.
Backend
FastAPI in Docker on a server you control, streaming responses over SSE and running the agents as real processes with tools.
Database
Postgres with authentication and row level security, so tenancy is enforced by the database and not by a filter someone remembered to add.
model providers
# the chat and the fleet can run on different providers
INTELLIGENCE_PROVIDER=openrouter
AGENT_FLEET_MODEL=claude-sonnet-5
# spend has a ceiling, per day, in the config
AGENT_FLEET_DAILY_CAP_USD=5
# writes are a separate switch from reads
AGENT_FLEET_EXECUTE_ENABLED=false
The honest part about cost
Every model call in the install runs on the owner API key. That is a deliberate design choice, and it has a consequence worth saying out loud before you buy: an implementation fee alone does not cover a system that thinks every day. The running cost is real, it is visible in the dashboard, and it is capped in the config.
A vendor who hides that line is quoting you a price they will have to walk back.
State of the system
What is on, what is off by default, what needs your keys
Written the way it actually ships. Anything that needs a credential you have not given stays dark rather than showing a number it invented.
- running
Multi-tenant install
organisations and projects under /o/:org/p/:project, with row level security in Supabase. One organisation runs on it today; two tenants have never shared an install, so isolation is still design and not result
- off by default
Fleet reads
the read tier agents run when you press Run, and write their findings into the account. The cadence on the roster is the one declared per agent: no line of code reads that column yet. Proven with one agent, not ten
- running
Intelligence with tools
proven end to end on a live install, including tool use through an OpenRouter model
- off by default
Fleet writes
off by default, and not a switch either: today the executor targets Google Ads, and pointing it at Meta is development. Action tier agents need explicit approval
- off by default
Scheduler
off by default. With the flag on, the loop runs one daily Google Ads review, not the fleet. Putting the fleet on a cadence is development, not configuration
- needs your keys
Ads and search dashboards
they light up with the client own credentials. On GA4, no account means a 503 instead of an invented number. The SEO screen is not there yet: with no cache it answers 200 with zeros, which is the next thing to fix
Day one
What switches on with nothing but the client own accounts
The test for every feature on this page is the same: does it work with an account the client already owns. If it needs anything else, it is not part of the install.
Creatives to campaign
A video the client already has becomes a Meta campaign, built and left paused for a human to approve.
needs: the client Meta ad account
Analytics that stops being manual
The monthly report becomes a screen instead of an afternoon of copy and paste.
needs: Meta, Google Ads and GA4 access
SEO and Keyword Scout
Keyword Scout, market demand and the forecast all read from Ahrefs, whose API is Enterprise only. This one breaks the rule of this page and is not sellable today.
needs: an Ahrefs subscription that is MINE, not the client account
Leads and funnels
A light CRM with funnel, products and event timeline, for the client who never adopted one.
needs: nothing external
Second Brain, agents, squads, skills
The operating memory of the account, with the fleet reading from it before it acts.
needs: a login and an owner
See it running on a real account
Thirty minutes, screen shared, on an install with live data. You bring one account you already own and we read it together. If the fit is wrong you will know inside the call.
Book a walkthrough