You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fastedit-mcp console entry point. The MCP server is now installable as a standalone binary alongside fastedit and fastedit-hook, so MCP config files can use "command": "fastedit-mcp" with no hardcoded venv paths, no python -m invocation, no python vs python3 ambiguity. Works identically in Claude Code, Cursor, Continue, and any other MCP host.
Fixed
fastedit pull silently skipping weight download. The cache-hit check used any(cached_path.iterdir()), which returned True on a partial cache (config.json + tokenizer.json present, .safetensors shards missing). fastedit pull then reported Model ready without downloading the weights, and every subsequent edit call failed at inference time with a cryptic model-load error. The check now requires at least one *.safetensors file in the cache before returning, so partial caches trigger a real download.
Changed
README install + MCP setup aligned with how the package actually ships. Primary install commands now include the mcp extra (uv tool install 'fastedits[mlx,mcp]' on Apple Silicon, [vllm,mcp] on GPU, [mcp] for external-server setups) — previously users who followed the README ended up with the MCP server importable but its fastmcp dependency missing, and got ModuleNotFoundError: No module named 'mcp' on first connect. fastedit pull examples now use the required --model mlx-8bit / --model bf16 flag (--model became required in 0.2.8). The MCP config example now uses "command": "fastedit-mcp" (the new entry point) instead of "command": "python", "args": ["-m", "fastedit.mcp_server"].