Self-hosted AI coding agent

What “local-first” actually means

Local-first does not mean “no network.” It means the orchestration loop: planning, tool dispatch, context compression, memory writes, and audit trails: executes in a process you run (TUI or web UI on localhost). Your repository, shell, git state, and optional Obsidian vault are accessed from your machine, not from a vendor sandbox you cannot inspect.

When you send a message, AgentHarness.send() binds chat context, runs intent classification on the fast model, then enters a multi-round ReAct loop. Each round can call dozens of tools; outputs land in tiered context (hot rounds verbatim, older output distilled to .agent_artifacts/). Session JSONL under .agent_sessions/ records the full trace for compliance or debugging.

What stays on your disk

• Session traces: .agent_sessions/<taskId>.jsonl when AGENT_SESSION_JSONL is on (default).
• Workspace artifacts and distilled tool output: .agent_artifacts/.
• Structured memory: .agent_notes.json with scoped notes (chat / workspace / global).
• Embedding index: ~/.liminal/memory.index.json (user-global, not trapped in a temp workspace).
• Optional Obsidian vault: real markdown files under Facts/, Entities/, Reflections/, etc.
• Runtime prefs (non-secret): .agent_runtime_prefs.json for harness toggles without editing .env.

What may still use the cloud

LLM completions go to whatever you configure in AGENT_API_BASE_URL. OpenRouter, OpenAI, Anthropic, xAI, or a local OpenAI-compatible endpoint. Embeddings for hybrid recall use AGENT_EMBED_MODEL when set; disable with an empty embed model for BM25-only.

Vision sidecar (vision_analyze) and optional safety judge are separate model calls you control via env. None of this requires a Liminal Cloud seat for Community Edition.

Approvals and destructive tools

run_shell and run_background are destructive: they go through the normal approval path unless AGENT_YOLO is enabled in a trusted environment. The safety judge (optional) can auto-clear obviously safe calls. Resource locks on file and shell prefixes prevent concurrent corruption.

Air-gapped and on-prem inference

Point AGENT_API_BASE_URL at Ollama or an internal gateway. The same tool surface (files, git, tests, memory) works; only model quality and speed change. See the Ollama guide for slug examples.

AGENT_API_BASE_URL=http://127.0.0.1:11434/v1
AGENT_MODEL=your-local-model
AGENT_FAST_MODEL=your-local-model

Who should run a self-hosted coding agent?

Teams with customer data in the repo, regulated industries, or strict IP policies often cannot paste code into a hosted IDE copilot. Freelancers and indie hackers choose self-hosted agents to swap models without re-learning a new product every quarter.

Liminal is fair-source (FSL-1.1-MIT), so you can read the harness, audit tool behavior, and keep JSONL session traces for compliance. That combination is what people search for when they look for a Cursor alternative that stays on their machine.

Self-hosted vs ChatGPT or Copilot

ChatGPT answers questions in a browser. GitHub Copilot completes lines inside an editor. Liminal runs a full tool loop on your PC: edit files, run tests, query git, browse docs, and write memory to disk. You bring any OpenAI-compatible model, including local inference.

For a side-by-side breakdown, see Liminal vs ChatGPT and Liminal vs GitHub Copilot on our compare hub.