· via dev.to (home feed)
Developer open-sources GitSeed, an Apache 2.0 pipeline linking local GPU models with cloud coding agents
A developer has released GitSeed under Apache 2.0, a self-hosted pipeline that routes coding work between local GPU models and cloud agents while keeping Git credentials away from the agents themselves.

A developer has open-sourced GitSeed, a self-hosted pipeline built to let local GPU models and cloud-based coding agents collaborate on the same work safely. According to the author's announcement on dev.to, the core platform is now available under an Apache 2.0 licence, with a three-step Docker Compose setup published in the project's GitHub repository.
A workflow split in two
The author, Trent Seed, describes a setup many developers will recognise: cloud subscriptions and APIs such as Claude Code, Codex and Gemini CLI on one side, and open-weight models served locally through Ollama or vLLM on the other, with multiple terminal windows running parallel git worktree loops.
The post identifies three recurring problems. First, coordinating local and cloud models was clumsy — the goal was to have stronger cloud models debate specifications and review diffs while local hardware ground through implementation, without plans being copied between tools by hand. Second, ideas that arrive away from the desk had no clean way to enter an autonomous pipeline, nor a safe way to check on it from a phone. Third, and most seriously, every local agent ran under the developer's operating-system user, which meant read access to the .ssh and .aws directories and to live Git push tokens.
What GitSeed does
The tool centres on units of work the author calls Seeds, which can be submitted from a command line, a web interface, or by voice through a smartwatch. Agents then draft a specification and run tests inside isolated containers. A human can approve a review gate from a phone, and step into an in-browser terminal or VS Code on the desktop when hands-on work is needed.
During ideation, planning and review, up to three models can run in parallel to provide second opinions before the specification is locked. Execution then targets one of three backends: cloud APIs reached through keys capped by LiteLLM, local GPUs with no traffic sent to external model APIs, or an opt-in single-user Personal Mode that runs against a provider subscription, subject to that provider's terms.
Credential isolation and audit receipts
The security design is the most distinctive part of the project. Every stage is split into three containers: prepare, agent and finalize. The agent container holds no Git credentials at all. Once the agent exits, the finalize container mounts the workspace read-only, blocks edits to protected paths, scans the diff for leaked secrets, performs the push from a clean .git directory, and signs an Ed25519 audit receipt — versioned as gitseed.receipt.v1 — onto the resulting pull request.
The pipeline also works on plain local folders before any GitHub App integration is configured. Merged changes fast-forward back into the developer's working tree without overwriting uncommitted edits, according to the post.
Stated limits
The author is unusually direct about what the tool does not solve. Container isolation does not stop an agent from writing buggy logic inside ordinary source files, so automated review and human diff inspection remain essential. Personal Mode also estimates spend from token telemetry rather than enforcing a hard dollar cutoff at a proxy, which means cost control is approximate rather than guaranteed.
It is worth noting this is a first-party announcement on a personal blog rather than independent coverage, and the claims about the tool's behaviour come from its own developer.
Why it matters
Hybrid AI coding setups — cloud frontier models paired with local open-weight models on personal GPUs — are increasingly common, but the glue between them is usually manual. GitSeed attempts to close that gap with two ideas worth watching even for people who never install it: giving coding agents zero Git credentials by construction, with a separate container doing all pushes and secret scanning; and signed audit receipts attached to pull requests so that automated changes leave a verifiable trail. For teams handling proprietary code, the local GPU mode that produces no external model API traffic is also a meaningful option, and the Apache 2.0 licence permits commercial adoption and modification.
- #open-source
- #ai-agents
- #developer-tools
- #llm
- #self-hosted