Rosetta

Centralized Cross-Application Context for AI Agents



Christopher R. Bilger

May 22nd, 2026

Agenda


  • Problem Statement
  • What Rosetta Is
  • Get It Running
  • What It Unlocks
  • Roadmap + Q&A

Context Decay Is a Tax on Every Multi-Repo Session


  • 11 services across as many repos
  • Each repo has its own conventions, test commands, deploy story
  • Documentation drifts the moment code changes
  • Agents either hallucinate or interrupt the engineer to re-ask

Cross-Service Wiring Is Where It Hurts Most


"Catch wiring drift during service-adds, not six months later."

Things We Tried First


  • CLAUDE.md / AGENTS.md in each repo - drift the moment code changes
  • Scattered READMEs - agents have to know to look at all of them
  • Pasting context into the chat by hand - burns engineer time, never consistent
  • Vector search over docs - retrieves text, misses the relationships between services

What We Needed



  • Single source of truth that agents can pull from on demand
  • Updated when the code updates, not when someone remembers
  • Repo-aware, not one-size-fits-all
  • Drop-in for Claude Code - no bespoke client

Rosetta in One Slide


  • An MCP server that ships as one npm package
  • Registered once at user scope - every Claude Code session picks it up
  • Serves bootstrap payloads tailored to the repo you're in
  • Two content streams: hand-curated skills + auto-synced API contracts

Three Flavors of Skills


Shared

Identical across every repo - HIPAA, auth, naming

Repo-Specific

How to test, deploy, run migrations here

Cross-Service

API contracts between this repo and the ones that call it


Same MCP server, three indexes - picked at bootstrap time based on which repo you opened.

The 5 Tools You'll See Most


Tool Purpose
bootstrap(repo) Session entry point - tailored payload
get_skill() Fetch full skill content
get_cross_repo_impact() Impact analysis when a service changes
get_map() Service registry + dependency graph
get_global_context() Org-wide standards (HIPAA, auth, naming)

5 More for Deeper Dives


Tool Purpose
list_skills() Search skills by repo/category
get_service_overview() Service README
search() Keyword search across all .md files
get_agent() Load specialized agent personas
get_domain() Cross-cutting domain knowledge (reserved)

Two One-Time Setup Steps

              
# 1. Authenticate with GitHub Packages
echo "//npm.pkg.github.com/:_authToken=YOUR_PAT" >> ~/.npmrc
echo "@talkiatry:registry=https://npm.pkg.github.com" >> ~/.npmrc

# 2. Register the MCP server (user scope, every session picks it up)
claude mcp add -s user rosetta -- npx -y @talkiatry/rosetta-mcp-server@latest
              
            

PAT needs read:packages - generate at GitHub → Settings → Developer settings → Personal access tokens.

Per-Repo: The Bootstrap Reminder Hook

              
// <repo>/.claude/settings.json -- Ensure that you update <repo> to the actual repository name
{
  "hooks": {
    "UserPromptSubmit": [{
      "matcher": "",
      "hooks": [{
        "type": "command",
        "command": "echo '[rosetta] REMINDER: Call bootstrap(repo: \"<repo>\") before proceeding.'"
      }]
    }]
  }
}
              
            

Watch It Work

From zero to a bootstrap'd session in ~90 seconds.

Cross-Repo Impact, Before You Write the Diff

  • Agent calls get_cross_repo_impact("ehr-service") before touching the API
  • Sees every downstream consumer in one shot
  • Catches breaking changes during draft, not in PR review (or later)

API Contracts That Don't Lie


Before

Agent guesses field names from outdated README → integration breaks at runtime

After

Agent calls get_skill("cross-service", "ehr-service", "patient-api") → contract is regenerated from current source on every sync

rosetta-sync: How It Stays Current


  • Update existing skills against fresh diffs
  • Discover skills the repo is missing
  • Mark stale when source no longer matches
  • Delete skills that no longer apply (cross-service only)

Sonnet first. Opus on retry. Keep the last good version if both fail.

"Why do I care?"


Less Drift

Context refreshed from source, not memory

Faster Onboarding

Every Claude Code session starts pre-briefed

Fewer Cross-Service Incidents

Contracts surfaced before code lands

HIPAA + Auth Defaults

Org standards delivered to every agent session


No hard numbers yet - early adopter feedback to come.

What's Next



  • domains/ namespace for cross-cutting knowledge (reserved, empty)
  • ignore_discoveries to mute noisy skill suggestions in sync-config.yml

Conclusion + Q&A


  • Problem Statement
  • What Rosetta Is
  • Get It Running
  • What It Unlocks
  • Roadmap + Q&A

Powered By



  1. A single, static webpage.
  2. reveal.js
  3. highlight.js
  4. Rosetta GitHub repo