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.
Portfolio: alexeiudall.com — career evidence, case studies, writing, projects, and a public read-only recruiter MCP.
Agent infrastructure. Six Model Context Protocol servers, all TypeScript, all built around one idea: an agent should be able to do the job without being handed the keys to everything.
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. 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.
| Server | What it is |
|---|---|
| credential-broker-mcp | Agents use gated APIs without ever holding the keys. Declarative policy, deny-by-default, secret injected at the last moment, content-free audit record. One decision function shared by the dry-run explainer and the live path, so what it reports and what it enforces cannot drift apart. |
| agent-handoff-mcp | Local-first bounded, auditable handoffs between AI agents. Lifecycle state machine, actor-scoped transitions, depth caps. |
| obsidian-github-mcp | Permissioned gateway letting agents propose changes to a GitHub-hosted Obsidian vault. Path policy, content policy, audit log. |
| jobscout-mcp | Privacy-first, bring-your-own-connections job discovery across multiple sources. Normalisation, deterministic deduplication, provenance. Stops at discovery: no CVs stored, no ranking, no applying. |
| source-pack-mcp | Structured research source packs. Facts, quotes, numbers, dates, primary links, and a coverage map showing which claims are consistent, contested, or isolated. |
| website-content-mcp | A site's content in agent-readable form. Real DOM parsing to clean markdown, sitemap discovery, disk cache, respects robots.txt, rate-limited. |
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 |
|---|---|
| Meta-Agent-OS | A Markdown operating system for diagnosing, designing, costing, and operating multi-agent systems. Upstream of code, not a runtime. Ten specialist personas, ordered stages, human decision gates. Most multi-agent projects fail because the team builds before it diagnoses. |
| agentops-template | Governed, repo-native operating layer for AI-assisted development. Cross-agent contracts, permission gates, durable state, persona councils, drift checks, and one canonical verification path. |
| obsidian-multi-brain | Local-first architecture for one canonical vault and governed downstream brains. Includes a validated mixed-authority bridge for public-safe agent context, bounded workspaces, and quarantined CRM proposals. |
| Prompt-Library | Curated system prompts for building with AI. Static site, no build step. |
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. The
MCP servers are Apache-2.0, and their tests run in CI. I would rather ship
something small that works than something broad that does not.
Currently building, as of August 2026: one governed agent stack around Control Room, CRM, the local-first vault, explicit handoff/credential/proposal boundaries, and a composable acquisition-to-evidence pipeline.
I reviewed all 74 repositories and moved 24 more into GitHub's recoverable archive state, taking the account from 7 to 31 archived repositories. Nothing was deleted. The active portfolio is now organized around 19 Core systems, 18 bounded product candidates, 10 time-boxed learning projects, and 27 strategic archive/low-maintenance holdings.
The consolidation rules are deliberate:
- Public MCP repositories own implementation; private deployments own policy, identities, configuration, and operations.
website-content-mcpowns bounded acquisition;source-pack-mcpowns evidence extraction, provenance, coverage, and persisted packs.Meta-Agent-OSowns multi-agent methodology;agentops-templateowns repository execution and verification.- The Obsidian vault remains canonical; satellites receive derived context and return proposals through governed paths.
This pass also merged validator-backed Pattern D support into
obsidian-multi-brain, production hardening into source-pack-mcp, npm
release preparation into obsidian-github-mcp, and the cross-platform
AgentOps template checks.
- Portfolio: alexeiudall.com
- LinkedIn: alexeiudall
- X: @Sarut0biSasuke