> great.sh

One MCP server, five AI backends. Alpha — open to testing and feedback.

great.sh is a single Rust binary that puts every AI CLI on your machine behind one MCP server — Gemini, Codex, Claude, Grok, and local Ollama — and can provision the rest of your AI dev environment while it's at it. Open source, Apache-2.0, built by Superstruct.

Quick start — the MCP bridge in 30 seconds

curl -sS https://great.sh/install.sh | sh          # install the great binary
claude mcp add --scope user great-bridge -- great mcp-bridge
claude

That's it. Claude Code now has bridge tools: ask it to "use great-bridge to ask gemini for a second opinion on this diff" or "run this refactor question past the local ollama model". The bridge auto-detects whichever backend CLIs are installed (gemini, codex, claude, grok, ollama) — no Node.js, no extra config.

Drop --scope user to register it for the current project only.

The MCP bridge

great mcp-bridge is a stdio JSON-RPC 2.0 MCP server that multiplexes AI CLI backends behind one tool surface:

                                        ┌──> gemini CLI   (GEMINI_API_KEY)
 Claude Code ──­ MCP (stdio) ──> great ──┼──> codex CLI    (OPENAI_API_KEY)
 (or any MCP client)          mcp-bridge├──> claude CLI   (login)
                                        ├──> grok CLI     (XAI_API_KEY)
                                        └──> ollama       (local, no key)

• Sync and async — prompt for a single round-trip; run/wait/ get_result/kill_task for long-running background tasks with timeouts and auto-cleanup
• Higher-level tools — research (query + file context) and analyze_code (review / explain / optimize / security / test), plus clink for spawning an isolated subagent with a custom system prompt
• Presets — minimal (1 tool), agent (7, default), research (9), full (10): expose only what the client needs
• Session resume — continue Claude/Codex conversations across calls
• Guardrails — optional allowed-dirs allowlist for file-reading tools, per-task timeouts, opt-out of auto-approval flags

Configure it in great.toml (all optional):

[mcp-bridge]
backends = ["gemini", "ollama"]   # default: auto-detect all installed
default-backend = "gemini"
preset = "agent"                  # minimal | agent | research | full
timeout-secs = 300
allowed-dirs = ["~/src"]          # restrict file-reading tools

Full reference — backends, config schema, adding a backend, security notes: docs/mcp-bridge.md.

Environment provisioning

The same binary takes a machine from blank to a fully configured AI dev environment:

• One command setup — great init then great apply: tools, MCP servers, agents, and secrets from a single great.toml
• MCP server management — install, configure, and credential-inject MCP servers from the registry
• Secret management — great vault sources API keys from env, 1Password, Bitwarden, or the macOS Keychain and injects them where tools expect them

Command	Description
great init	Initialize a new great.sh environment
great apply	Apply configuration to the current environment
great status	Show environment status
great mcp-bridge	Run the multi-backend MCP bridge server
great mcp	Manage MCP servers
great vault	Manage credentials via local secret providers
great doctor	Diagnose environment issues
great update	Update great.sh to the latest version
great diff	Show configuration diff
great template	Manage configuration templates
great sync	Export/import config snapshots (local storage)
great loop	Install and manage the great.sh Loop plugin

The great.sh Loop

An evidence-gated AI agent methodology shipped as a Claude Code plugin: four functional roles instead of a fixed pipeline, recalibrated for models that plan and verify their own work.

• builder — implements the spec, runs quality gates, answers findings with evidence
• verifier — adversarial: tries to prove the change broken or insecure (correctness, regression, security, performance)
• reviewer — read-only quality review (structure, simplification, UX, output design, docs)
• scout (optional) — read-only recon for large or unfamiliar change surfaces

Plan (lead) → Build & Verify (builder + verifier + reviewer) → Finish (lead)

Your session is the team lead: it writes the spec, self-reviews it for gaps, spawns the team, and commits. A phase ends when its exit criteria are met — quality gates green plus no CONFIRMED blocking findings — never after a fixed number of rounds. Verifier findings must cite reproductions; the builder responds with rerun evidence, not re-argument.

Roles inherit your session model by default. Pin a tier per role in ~/.claude/teams/loop/config.json when the work demands it — e.g. Opus for security-audit-heavy verification, since Fable-class cyber safety classifiers can refuse security-probing work mid-audit.

great loop install            # registers the plugin with Claude Code
great loop install --project  # adds .tasks/ working state to the current repo
claude                        # then: /great:backlog → /great:loop

Templates

Pre-configured environment templates from architecton.ai:

• AI Full Stack (TypeScript) — Claude Code + Codex + Gemini, Playwright MCP server, gh CLI
• AI Full Stack (Python) — Python with uv, PostgreSQL MCP server, full AI agent setup
• AI Data Science — CUDA, Jupyter, Gemini for data analysis, database MCP servers
• AI DevOps — Terraform, AWS CLI, Docker, Kubernetes MCP servers
• AI Minimal — Just Claude Code with Filesystem and Memory MCP servers

Development

# CLI
cargo build
cargo test
cargo clippy

# Marketing site
pnpm dev
pnpm build:site

Status and roadmap

Alpha (v0.1.0). The MCP bridge, provisioning (init/apply/doctor), vault providers, and the Loop plugin work today. We welcome bug reports and feedback.

Being candid about scope: great sync currently exports/imports configuration snapshots to local storage. A hosted cloud-sync service is not planned — use the local secret providers (env, 1Password, Bitwarden, Keychain) to share credentials across machines.

• GitHub Issues
• Discussions

License

Apache-2.0 — see LICENSE

Built by Superstruct. Copyright 2025 Superstruct Ltd