ai-workspace

Bootstraps a dev machine to work with several AI coding CLIs, so work can move to another provider when tokens run low, ending in a fully local fallback. Everything is pinned in this repo; the host only needs mise, direnv, and op (1Password CLI).

Layout

One folder per provider (claude/, antigravity/, opencode/, ollama/), each with its own justfile surfaced from the root via mod . Every provider module exposes the same four recipes:

MCP servers are defined once, in mcp/servers.toml, and rendered into each CLI's native format by scripts/sync-mcp.py (just sync-mcp). Never hand-edit .mcp.json, .agents/mcp_config.json, or opencode.json — they're generated and gitignored.

Shared instructions live in AGENTS.md — opencode and Antigravity read it natively. Claude Code does not, so CLAUDE.md starts with @AGENTS.md and appends Claude-only notes below it.

Shared memory across every CLI is basic-memory — plain Markdown under memory/, committed like any other file, registered as the ai-workspace project by just doctor.

Secrets

secrets.env holds only references (safe to commit); each start recipe resolves them at launch via op run --env-file=secrets.env -- `, which injects the real values into that one child process's environment and never writes them to disk.

op run itself still needs to authenticate to 1Password to resolve those references. Once per terminal session, run: