One protocol, 18 servers, 270+ tools. An AI session that can resolve a repo, commit code, render a Confluence page, dispatch ten parallel agents, and measure its own cost — all through typed composition. No integration code. No shared state. Just tools that produce output other tools consume.
Servers don't import each other. They produce typed output; other servers consume it. The agent is the integration layer — it calls one tool, reads the result, passes it to the next. Rosetta names every handoff point so the wiring is discoverable, not memorized. Three chains illustrate the pattern.
Five tiers, strict dependency direction. Each tier consumes the one below it and knows nothing about the tiers above. External services sit at the top — consumed by domain tooling, never by platform.
The ground everything stands on. Context resolves paths, git-ops guards commits, opencode-config manages the runtime. One builder server and two shared libraries make the ecosystem self-extending — the nineteenth server is three commands away.
The single source of truth for all filesystem paths. Activate a context and every tool knows where to look. Nothing touches disk without going through context first.
Git with guardrails. Every commit is categorized and lint-checked. Amend, fixup, autosquash rebase — the sharp tools are there, but they verify your work first.
Configuration as code for the AI runtime. Which servers are active, which agents exist, which permissions apply. Export a setup, import it on another machine.
Scaffold servers, plugins, and commands. Drop a .ts file in src/tools/ and it's a tool — no registration, no routing.
startServer() to boot, auto-discover tools, wire transport. Type-safe git and glab operations. The foundation under every server on this page.
A pipeline that goes from color decision to published page in four tool calls. See Chromatic Language for the color system, Diagram Language for SVG rendering.
Full Confluence lifecycle — fetch, create, update, publish, search, screenshot, label, comment. Content goes from disk to wiki in one call.
The rendering engine. YAML specs in, Confluence XHTML out. Cards, tables, heroes, flow chains, badges, headings, TOC. Decompile↔compile for round-trip editing.
Describe structure, get geometry. Pipelines, topologies, state machines, swim lanes, sequences, trees, cycles, fan-outs — all auto-laid-out to SVG. Every diagram on this page was generated by this server.
21 hues × 11 shades, perceptually uniform via OKLCH. Named schemes, contrast presets, surface palettes for dark rendering. One resolve call returns every color decision for a page.
YAML page specs to standalone HTML5. The same spec format that drives Confluence pages, rendered to a static site with entrance animations, responsive layout, and dark theme. This page is a gutenberg render.
What the AI pays attention to, how it fans out work, how sessions avoid stepping on each other, and what it all costs.
Cognitive channeling. YAML flow definitions narrow which tools are available per phase and direct attention through signals. Definitions extracted from production, not designed upfront.
Parallel dispatch with dependency graphs. Plan missions, fan them out to subagents, fan results back in. Missions self-report completion and surface blocking questions.
The type system between servers. Every integration seam gets a name — ContextHint, RepoCwd, XhtmlDocument. Rosetta tracks who produces each type, who consumes it.
Multi-session awareness. See what other sessions are doing, claim paths you plan to touch, check for conflicts before editing. Soft locks, not hard blocks.
Session analytics. Which servers earn their keep? Which sessions cost the most? Daily spend, token usage, failure rates, session learnings.
Domain-specific servers for billing engineering. Disabled by default — enable when the work needs them.
Go build toolchain for subscriptions-api. Tests with filtering, lint, codegen, VCR cassette management.
Read-only BigQuery access. Discover schemas, search tables, run queries against billing data.
Log pattern analysis. Discover logging processes, build KQL queries, extract recurring patterns.
Cloudflare's internal gateway and third-party APIs, consumed as MCP servers.
Cloudflare's internal MCP gateway. 7 upstream services behind one connection: Backstage (17), Google Workspace (26), Jira (9), Sentry (7), Elasticsearch (6), Confluence (6), Docs (2).
Official Stripe MCP server for payment API operations. Runs via bunx — no local install needed.
Every diagram on this page was generated by the ecosystem it documents. The topology was rendered by svg-tools, themed by chromata, compiled by gutenberg, and deployed by a single tool call. That's the point — the system describes itself.