Remote MCP server
MCP Session Bridge
A shared memory layer for multi-model AI work: one durable place where different assistants can save, read, and continue the same conversation without losing the thread.
Live session package
One transcript, many assistants
USER
Explain the deployment constraints.
CLAUDE
Reads chunks 1-3, proposes a rollout.
GPT
Continues from the saved exchange.
Problem
AI assistants are powerful, but their memory often stops at the boundary of one client or one chat.
Approach
A narrow MCP server stores conversation history and exposes a predictable protocol for reading it back.
Recruiter signal
The project shows backend design, AI tooling fluency, security awareness, and operational discipline.
How it works
A durable bridge between assistants, not another chat UI.
AI clients
Claude, ChatGPT, Codex
Remote MCP
Streamable HTTP, FastMCP, Uvicorn
Session core
SQLite, transcript chunks, summaries
Audit layer
Offline viewer and protected admin UI
Model protocol
Designed around the way assistants actually work.
The bridge turns session continuity into a repeatable tool sequence. Assistants establish the session, inspect the overview, read every transcript chunk, answer with context, then save the exchange before the user sees it.
Create
A client creates or selects a durable session before work starts.
Read
The model fetches bounded transcript chunks instead of one oversized result.
Answer
The assistant responds with the full context of prior model exchanges.
Save
The latest user/model exchange is stored for the next assistant.
Summarize
Markdown summaries can be saved when a compact handoff is useful.
Engineering proof
Practical backend judgment in a small package.
Protocol fit
Built as a remote MCP server over streamable HTTP, with a tool flow models can reliably follow.
Security posture
OAuth authorization-code + PKCE, dynamic client registration, hashed bearer tokens, CSRF-protected admin changes.
Operational shape
SQLite persistence, Uvicorn deployment, health checks, Caddy/systemd templates, and local demo scripts.
Product restraint
The bridge stores conversation history only; it intentionally avoids pretending to manage project files.
Stack