Free, and open source.

DGC is released under the Apache License 2.0. You may use, modify and redistribute it, including commercially and inside a company, with no separate agreement and no fee. The license also grants the patent rights the contributors hold in the work. Its two obligations are the ordinary ones: keep the license and copyright notice with copies, and state significant changes you make. The repository’s LICENSE file controls.

DGC sends conversation context to the model route you select. With a model on the same machine, prompts, relevant code excerpts, and tool results can stay there. With a model on another host, a cloud API, or a subscription, that service receives the context needed to answer under its own terms. Optional web search, fetch, MCP, and other configured integrations receive the queries or tool arguments you direct to them. A configured language server receives workspace metadata and the full text of documents queried through code intelligence. Interactive DGC may also make a small version check that contains no prompt, code, or session content; a successful check is cached for a day, while a failed check may retry on a later launch.

Yes, the coding loop works without internet access when DGC is already installed and its model, toolchains, and any integrations you need are reachable locally. Cloud providers, subscription routes, web search, remote MCP servers, updates, and downloads naturally require a connection. The background CLI version check fails quietly when the site is unreachable and does not block the agent.

DGC requires Python 3.10 or newer. The one-line Bash installer supports Linux and macOS. On Windows, use WSL for the CLI; the VS Code or Cursor extension can connect to that workspace. OS confinement currently has truthful backends on Linux and macOS. A requested sandbox fails closed on an unsupported platform rather than pretending to isolate commands.

The standard installer puts the application in ~/.local/share/dgc and a symlink at ~/.local/bin/dgc; it does not require root. Remove those two paths to remove the installed program (an install made by an older installer may also have left ~/dgc). CLI-managed configuration, credentials, sessions, plans, checkpoints, and other private state live separately under ~/.dgc; remove that directory only if you also want to erase that CLI state. Delete the DGC extension through your editor’s Extensions view. Credentials saved by the extension live in the editor's SecretStorage, while extension settings and update metadata live in editor-managed storage; use the editor or operating system's credential and profile controls if you also want those records removed.

Each selected session becomes one JSONL record with an OpenAI-style messages array, portable assistant tool calls and tool results, plus small metadata such as model, project basename, turn and tool counts, edit counts, and an outcome flag. Reasoning traces, provider continuation blobs, images, and internal bookkeeping are omitted. Every emitted field is deep-scrubbed for configured secrets and high-confidence credential patterns. Export is read-only and local; DGC never uploads the file.

No product-usage telemetry is sent by the CLI or editor. DGC keeps sessions and operating metrics locally so features such as resume, rewind, and benchmark accounting can work. Interactive CLI launches may check vibedgc.com/version.json; successful checks are cached for a day and failed checks may retry later. A self-hosted VSIX checks /vscode/version.json at most daily by default, while Marketplace/Open VSX builds skip that request and the setting can be disabled. Neither request contains prompts, code, tool arguments, or credentials. Cloudflare's website-delivery processing and optional aggregate Web Analytics are described separately in the Privacy notice; DGC does not operate a custom website event endpoint or event dataset.

The current measured slice records the same model alias, caller-declared digest, endpoint URL, and task set for every agent; the saved runner verified the 32,768-token context size but model weights were not independently verified. Each exercise starts in an isolated workspace and is graded by its official test suite. The first round asks the harness to implement and test the stub; only a failure continues the same session for the final recovery round with a bounded failure tail whose paths are normalized to the live workspace. The result is pass@2. DGC’s stricter publication gate also requires complete engines, tasks, runner provenance, transport, and usage accounting; this 90-task slice is disclosed separately because it does not clear that gate.