Seventeen years selling, and I ship the infrastructure too. I work at the front of the business, in revenue and partnerships, and I build the tooling that makes that work repeatable. Most of what is here sits at the join between the two: agent infrastructure, knowledge systems, and the small tools that come out of doing the job.
Currently Head of Marketing at Teneo Protocol, a decentralised AI agent network.
- mcprack.dev: every MCP server I have published on one board, with install commands, plus the servers I actually run alongside them.
- alexeiudall.com: career evidence, case studies, writing, projects, and a public read-only recruiter MCP.
Agent infrastructure. Model Context Protocol servers, mostly TypeScript, all built around one idea: an agent should be able to do the job without being handed the keys to everything.
Hosting and quality around them. A gateway that puts auth, plans and rate limits in front of an MCP, and a shared harness that blocks a release if a server stops honouring its published contract.
Knowledge systems. Obsidian-centred vaults that agents can read from and propose into, without being able to quietly rewrite the source of truth.
Games. Small ones, mostly to learn. Phaser, Godot, pygame, and an older Unity one.
Each one solves a boundary problem: what the agent is allowed to reach, what it is allowed to see, and what it leaves behind. The full board, with install commands and registry status, is at mcprack.dev.
What "boundary" means in practice. This is the whole of an agent's authority in credential-broker-mcp, and there is no way to exceed it:
agents:
- id: researcher
credentials: [github-readonly] # the token itself never reaches the agent
methods: [GET]
allow: ["/repos/**"]
deny: ["/repos/*/*/keys"] # deny always beats allow
max_response_bytes: 262144A request must clear every gate: identity enabled, credential granted, method granted, URL inside the credential's own base URL, no deny match, and an explicit allow match. Redirects are refused outright, because a redirect could point anywhere and the injected credential would follow it.
| Project | What it is |
|---|---|
| mcp-host-gateway | Hosted MCP edge. Bearer JWT auth, plan-based entitlement down to the tool, prompt and resource, per-identity rate limits, streaming relay, and an audit line per call with no bodies or tokens. Fails closed at startup if anything in its config does not add up. |
| mcp-eval-harness | Shared contract tests for MCP servers. Golden fixtures and JSON Schemas, run through the official client, with a CI exit code that gates publishing. One question: does this server still honour the contract we published? |
| mcp-dashboard | Cost against use for your MCP toolbox across Claude Code, Codex, Gemini CLI and Cursor: RAM, context tokens, startup time and real call counts, with working on/off switches. On PyPI. |
| mcp-rack | The source for mcprack.dev. Static, no build step, one data file. |
I write the specification before the code, and I keep an AGENTS.md in every repository so that both humans and agents know the rules of the project. Public repositories carry an open licence, and their tests run in CI. I would rather ship something small that works than something broad that does not.
Currently building, as of October 2026: the layer that turns these servers into something other people can run and pay for. A hosted gateway in front of them, a shared contract harness gating every release, a governed read-only Postgres server for small-business installs, and a local control plane that tracks project stages, human decision gates and inference costs across every repository, with agents reporting into it over MCP.
- MCP servers: mcprack.dev
- Portfolio: alexeiudall.com
- LinkedIn: alexeiudall
- X: @Sarut0biSasuke