{"slug": "i-built-a-self-hosted-pipeline-to-make-my-local-gpus-and-cloud-coding-agents-it", "title": "I built a self-hosted pipeline to make my local GPUs and cloud coding agents work together safely — just open-sourced it (Apache 2.0)", "summary": "A developer named Trent Seed open-sourced GitSeed under Apache 2.0, a self-hosted pipeline that coordinates cloud coding agents (Claude Code, Codex, Gemini CLI) with local open-weight models running on Ollama or vLLM. The system runs up to three models in parallel for spec and diff review, executes each stage in prepare/agent/finalize containers so agents hold no Git credentials, and signs an Ed25519 audit receipt onto pull requests. Seed acknowledged limits: container isolation does not catch buggy logic in normal source files, and Personal Mode estimates spend from token telemetry rather than enforcing a dollar cutoff.", "body_md": "My AI coding setup was split in two: cloud subscriptions and APIs (Claude Code, Codex, Gemini CLI) on one side, local open-weight models (Ollama / vLLM) on the other, and four terminal windows running parallel git worktree loops on my machine.\n\nThree things kept bugging me:\n\nGetting local and cloud models to collaborate was clunky. I wanted frontier models debating specs and reviewing diffs, and fast or local GPU models grinding through implementation — without manually copying plans between tools.\n\nIdeas don't only happen at the keyboard. Half my best product fixes hit me away from my desk, with no clean way to feed an idea into an autonomous pipeline and check on it safely from my phone.\n\nHost security was an accident waiting to happen. Every local agent was running under my OS user with read access to ~/.ssh, ~/.aws, and live Git push tokens.\n\nSo I built ([GitSeed](https://gitseed.dev) · [https://github.com/gitseed-dev/gitseed](https://github.com/gitseed-dev/gitseed)) — and today I'm open-sourcing the core platform under Apache 2.0.\n\nWhat changed in my daily workflow\n\nCapture anywhere, pick up anywhere: I can toss a rough Seed into the pipeline from the CLI, web, or even voice on my watch, let agents draft the spec and run tests in isolated containers, approve a review gate from my phone, and drop into an in-browser terminal or VS Code on desktop when something needs a human touch.\n\nFrontier brains + local muscle: Up to 3 models run in parallel for Second Opinions during Ideate, Plan, or Review before locking the spec. Then Execute runs against LiteLLM-capped API keys, local GPUs (with no external model API traffic), or opt-in single-user Personal Mode (subject to your provider's subscription terms).\n\nAgents hold zero Git credentials: Every stage splits into prepare → agent → finalize containers. finalize mounts /workspace read-only after the agent exits, blocks edits to protected paths, scans diffs for leaked secrets, pushes from a clean .git directory, and signs an Ed25519 audit receipt (gitseed.receipt.v1) onto the PR.\n\nWorks on Local Folders too: Point it at a local directory on your laptop before setting up any GitHub App — merged changes fast-forward back into your working tree without clobbering uncommitted edits.\n\nHonest limits\n\nContainer isolation doesn't stop an agent from writing buggy logic inside normal source files (Review + human diff review are still essential), and Personal Mode estimates spend from token telemetry rather than enforcing a proxy dollar cutoff.\n\nIf you're mixing local GPUs and cloud coding agents on your own projects, I'd love to hear what your workflow looks like — repo and 3-step Docker Compose setup are at [https://github.com/gitseed-dev/gitseed](https://github.com/gitseed-dev/gitseed).\n\nTrent Seed", "url": "https://wpnews.pro/news/i-built-a-self-hosted-pipeline-to-make-my-local-gpus-and-cloud-coding-agents-it", "canonical_source": "https://dev.to/trent_seed_0ad49423e2f31f/i-built-a-self-hosted-pipeline-to-make-my-local-gpus-and-cloud-coding-agents-work-together-safely--4bh5", "published_at": "2026-11-18 03:01:46+00:00", "updated_at": "2026-11-18 03:11:28.594967+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "ai-infrastructure", "ai-safety"], "entities": ["GitSeed", "Trent Seed", "Claude Code", "Codex", "Gemini CLI", "Ollama", "vLLM", "LiteLLM"], "also_reported_by": [], "takeaway": "GitSeed splits each agent stage into prepare, agent and finalize containers, mounting /workspace read-only after the agent exits and pushing from a clean .git directory so coding agents never hold Git credentials.", "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/i-built-a-self-hosted-pipeline-to-make-my-local-gpus-and-cloud-coding-agents-it", "markdown": "https://wpnews.pro/news/i-built-a-self-hosted-pipeline-to-make-my-local-gpus-and-cloud-coding-agents-it.md", "text": "https://wpnews.pro/news/i-built-a-self-hosted-pipeline-to-make-my-local-gpus-and-cloud-coding-agents-it.txt", "jsonld": "https://wpnews.pro/news/i-built-a-self-hosted-pipeline-to-make-my-local-gpus-and-cloud-coding-agents-it.jsonld"}}