Codex AnalyticaService · 03 of 03
We build the intelligence layer that makes everything else faster: MCP servers exposing your stack to Claude, GPT, and Gemini, agentic workflows that run on cron, and reports that write themselves.
What we deliver
Native Model Context Protocol servers exposing your APIs and data to any LLM client: Claude Desktop, Cursor, Continue, your own agent.
Multi-step Claude / GPT workflows that run on cron, react to events, and write back to your stack: durable, retryable, cost-controlled.
Reports that write themselves (Friday revenue summaries, weekly board notes, anomaly call-outs), grounded in your real data, not hallucinated.
+61 401 ··· 928
AU · agent-pickup 1.4 s
Real-time voice agents that pick up the phone, qualify a lead, book a call, and write the transcript to your CRM.
Methodology
Same Friday demos, same documented handover. We instrument every prompt, every tool call, every token, so you know exactly what the agent did.
Where AI adds genuine value vs where it's a gimmick. Use-case mapping with a written ROI model. We've turned down work that didn't pencil.
no-AI mode available
Honest
Prompt architecture, tool surface, evaluation harness, fallback paths, cost ceilings. All in writing before any LLM call goes live.
before deploy
Eval-first
Iterative builds with weekly accuracy checks against your data. Cost dashboards from day one. Production rollout behind a feature flag.
token budgets
Per-step
Prompt registry, eval traces, runbook for failures, retraining plan. Your team can extend the system without us.
of every run
Full trace
Tech stack
We don’t pretend to be neutral. We pick stacks we’ve put on a SOC-audited tenant.
Models
Protocols
Agents & workflow
Storage & retrieval
Sample deliverable
A Friday-revenue-report agent that pulls from Power BI + Stripe, drafts a 200-word narrative, validates against source data, and posts to #ops with citations. Runs every Friday at 9am AEDT.
Default is Claude (Anthropic) for reasoning + tool use, GPT-5 for structured output workloads, Gemini for cost-sensitive batch. We benchmark before committing, then route through Vercel AI Gateway so model choice is a config flag, not a rebuild.
Three things: structured output with JSON schema validation, source-data citation on every claim, and human approval gates for high-stakes outputs. We never deploy AI output to a user-facing surface without an eval harness running on a real-data sample.
Default is EU-region inference (Anthropic EU + OpenAI EU). For zero-retention workloads we route through enterprise endpoints with ZDR enabled. We can run no-LLM mode if your data class doesn't allow third-party processing.
No. AI eats the tedious parts (pulling data, formatting reports, drafting narratives) so analysts focus on insight and decision support. The teams who deploy AI well end up needing more analysts, not fewer.
Yes. We run a dozen in production, including the public Qlik MCP and Wiki MCP. The protocol is stable, the SDKs are mature, and Anthropic / OpenAI / Vercel all ship native clients. You're not betting on a science project.
AI & Automation
We'll come back with a fixed-price agent that ends it.
A 30-minute discovery call to scope your project. We'll tell you honestly whether we're the right fit, and if we are, we'll have a proposal within a week.