2026: AI employees and MCP connectors
The Forward Deployed Engineer work, where AI systems hold real jobs inside a business and nine live MCP connectors run an agency's operations.
- Status
- stable
- Trust
- human-reviewedhuman:rrizwan98 ·
- Approved
- human:rrizwan98 ·
Today the work is AI employees, systems that hold a real job inside a business, deployed where the business already lives: WhatsApp, Gmail, the phone, and the AI tools a team already uses. 20 repositories by the end of August, 11 of them private.1
What happened
- AI employees. A task-execution employee and a manager agent that works a Gmail inbox (reads mail, answers questions, produces research and reports) in January; a backend-developer employee that writes and ships backend code the same month; WhatsApp-based employees in February and March, one of them a WhatsApp assistant that builds custom AI agents, with container isolation and agent swarms.2
- Cold-outreach automation. An agent pipeline that runs cold email campaigns, in February.
- A voice agent. Real-time, phone-style conversation on the OpenAI Realtime API, in March.
- Self-hosted infrastructure. A Honcho memory server on Hugging Face Spaces backed by Postgres, in May.
- A marketing connector. A remote MCP server carrying 45 packaged marketing skills and their orchestration logic, in June.
- Agency systems. A creative service, a lead pipeline, a client website with an AI assistant, and an integration bridge, all private.
The live MCP connectors
An MCP connector is a remote server that ChatGPT, Claude, Codex or Claude Code can attach to. Each one packages a company's process and rules, so any AI that connects behaves like a trained employee of that company. Nine were live and answering on 26 August 2026:3
| Connector | What it does | What it refuses |
|---|---|---|
| canz-proposal | Writes client proposals the agency's way, one section at a time | Whole proposals dumped in one call |
| agent-factory | Builds AI agents under an operating manual: plan and specs first, tests first | Code before a plan and acceptance criteria exist |
| agentic-developer | Turns a rough development request into a researched plan, tested changes and a delivery package | Skipping the plan |
| canz-website | Designs, builds and QA-checks websites from a plain-language brief; seven design documents come before any code | Placeholder copy, secrets, a hero section scoring under 12/12 |
| canz-website-qa | Audits websites and scores them; the server recomputes every score | Inflating its own grades |
| canz-ads | B2B ads strategy and copy | Writing ads before nine campaign inputs are confirmed |
| canz-research | Market research reports with evidence floors | Reports without sources, vague quantities |
| canz-sor | The agency's read-only record of how work is done | Writes of any kind |
| canz-pm | Read-only reporting over the agency's client board | Writes, and lists the data cannot support |
Live connectors are named and described by what they do and refuse. How any of them is built is not published here.
The AI-native agency
In July 2026 a marketing agency's delivery moved onto an AI-native operating model: work enters from the CRM, AI staff do the job, one person reviews and approves. Names, counts and architecture are deliberately outside this record.
The public repositories
| Created | Repository | What it is |
|---|---|---|
| 2026-02-21 | calculator-agent | Calculator agent with a natural-language interface, OpenAI Agents SDK |
| 2026-02-25 | calculator-agent-new | Natural-language calculator agent, OpenAI Agents SDK |
| 2026-02-26 | ai-employee-nanoclaw | WhatsApp-based AI assistant that builds custom AI agents, in four phases |
| 2026-02-26 | basic-calculator-agent | Calculator agent with a Next.js front end and FastAPI back end using ChatKit |
| 2026-02-28 | powerfit-fitness-coach | AI fitness coaching platform: workout plans, nutrition tracking, exercise library |
| 2026-03-28 | (name withheld) | Voice agent on the OpenAI Realtime API; the repository carries an agent's name, which this record does not publish |
| 2026-04-04 | ai-agency | Landing page for a custom AI agents and AI employee service |
| 2026-05-07 | honcho-self-hosted | Self-hosted Honcho memory server on Hugging Face Spaces with an external Postgres |
| 2026-06-30 | lara-marketing-connector | Remote MCP server with 45 marketing skills and orchestration logic for any MCP host |
Private work
Eleven private repositories: a family of AI-employee builds (a task-execution employee, a manager agent on Gmail, a backend-developer employee, two WhatsApp-based employees), a cold-outreach automation, a co-founder experiment, and four agency systems for a marketing company (a creative service, a lead pipeline, a client website with an AI assistant, an integration bridge).1 They are counted here and never named.
The pattern behind all of them
The server enforces the process. These systems do not rely on the AI behaving. The pipeline rejects skipped steps, unverified claims and invented facts, which is what makes them safe to hand to a business.2
What this era is teaching
Real deployment is 20% model and 80% process, integration and accountability. That is the actual job of a Forward Deployed Engineer.2
Back to the journey overview, or on to the journey in numbers.
Footnotes
-
GitHub API, repositories owned by rrizwan98, read 26 August 2026: 20 repositories created between 1 January and 26 August 2026 (9 public, 11 private). ↩ ↩2
-
From Predictive AI to Forward Deployed Engineer: My 6-Year Journey, Raza Rizwan, July 2026. ↩ ↩2 ↩3
-
MCP initialize probe, 26 August 2026 19:36 UTC: canz-proposal, agent-factory, agentic-developer, canz-website, canz-website-qa, canz-ads, canz-research, canz-sor, canz-pm, all HTTP 200. ↩
Sources
- [^github-repos]GitHub API, repositories owned by rrizwan98, read 26 August 2026
- [^journey-note]From Predictive AI to Forward Deployed Engineer: My 6-Year Journey, Raza Rizwan, July 2026
- [^connector-probe]MCP initialize probe of the nine connector endpoints, 26 August 2026, 19:36 UTC, all answering HTTP 200