Choose what your assistant knows.

Golden Gate tells the AI assistant you chat with what you have been building in Claude Code, Codex and OpenCode. It sends redacted summaries of the projects you pick to your own Golden Gate Bridge server, and nothing else.

QuickstartHow it stays private

Every project, and what happens to it.

Open Golden Gate in your terminal with goldengate. Find a project, share it with space, read exactly what would leave with p, then review and send with s.

The Golden Gate TUI: moving through the project list, searching for vihayas, sharing it, previewing its summary, then reviewing and sending a sync until every shared project shows synced.
Recorded from the real TUI against a local server, with synthetic sessions.
  • [x]Shared
  • [ ]Not shared
  • [~]Some tools shared
  • [-]Denied
  • ▲ 33 sessions to sync

Two programs, one direction.

Golden Gate runs on your laptop and Golden Gate Bridge runs on your server. Data only moves from one to the other: the bridge never reaches into your machine, and every payload leaves through one function that checks it first.

  1. 01

    Golden Gate

    On your laptop

    Reads your coding tools' history, read-only. Only projects you picked go further.

  2. 02

    Redact and summarize

    Still on your laptop

    Secrets and paths are scrubbed before and after a local summary.

  3. 03

    Golden Gate Bridge

    On your server

    A small server you run. No hosted service, no telemetry.

  4. 04

    Your assistant

    Gemini, ChatGPT, Claude

    Reads over MCP with a read-only token you can revoke.

The preview is the payload.

Before anything leaves, you can read it. goldengate sync --dry-run prints the same thing.

Paranoid mode in the preview: the exact JSON of a session payload that would be sent, with source, project, level and a summary of goals, outcomes and messages.
Paranoid mode (P) prints the exact bytes of every payload that would leave. Titles, goals, outcomes, file names and command names. No transcript, no tool output, no file contents.

What it is for.

Your assistant can answer questions about your own recent work, because it can read the summaries on your Golden Gate Bridge server.

A weekly review in one question

Ask what you worked on this week and get an answer across every project and every coding tool.

Pick up where you left off

The goals, recent sessions and open threads of a project, without scrolling old chats.

Find what is still open

Pull requests opened but never confirmed merged, and commits that were only mentioned.

Find the session you half remember

Search your past work by what it was about, then open that one session.

See all use cases, or the nine MCP tools the assistant reads with.

Nothing is shared until you pick it.

Nothing leaves until you pick it

New projects stay on your machine, and a deny rule always beats an allow rule.

Redacted twice

Secrets and paths are scrubbed before and after summarizing. Best effort, so the preview is always there to read.

A token that can only read

The assistant's token can neither write nor delete, and you can rotate or revoke it at any time.

Read the security model: tokens, the opt-in model summarizer, prompt injection, and what it cannot protect against.

About ten minutes, start to finish.

There are no packaged releases yet. Build both programs, Golden Gate (goldengate) and Golden Gate Bridge (goldengatebridge), with Go 1.26 or later:

go install github.com/divyanshusingh2903/goldengate/cmd/...@latest
  1. 1

    Run Golden Gate Bridge behind HTTPS

    export GOLDENGATE_ADMIN_TOKEN="$(openssl rand -base64 32)"
goldengatebridge serve --addr :8080 --behind-tls-proxy --db /data/bridge.db
  2. 2

    Connect Golden Gate on your laptop

    goldengate init --server https://bridge.example.com
  3. 3

    Pick a project, look, then send

    goldengate share add ~/code/myproject
goldengate sync --dry-run
goldengate sync
  4. 4

    Point your assistant at it

    https://bridge.example.com/mcp
Authorization: Bearer <read token>

Your assistant only supports OAuth? Start the server with --oauth --public-url and approve the connection with your admin secret. The README has every option.

Sources

  • CC
    Claude Code~/.claude/projects/**.jsonl
  • CX
    Codex CLI and desktop~/.codex/sessions/**/rollout-*.jsonl
  • OC
    OpenCodeits SQLite database, read-only
  • AI
    claude.ai exportgoldengate import-claude-ai export.zip
  • GPT
    ChatGPT exportgoldengate import-chatgpt export.zip

Roadmap

  1. Done
    M1 CoreAdapters, local summarizer, redactor, selection rules, dry-run sync
  2. Done
    M2 ServerSQLite store, ingest API, scoped tokens, pairing, MCP over streamable HTTP, audit log
  3. Done
    M3 More sourcesCodex, OpenCode, claude.ai and ChatGPT exports, one project across tools
  4. Done
    M4 TUISetup wizard, project picker, exact preview, two-step sync, server and settings screens
  5. Next
    M5 DeployPostgres and Firestore stores, Dockerfile, Cloud Run script, docker compose with Caddy
  6. Later
    M6 ReleaseSigned builds for Linux, macOS and Windows, SBOM, Homebrew tap, finished hardening checklist

The full plan is in DESIGN.md.