openheim

Openheim is an open-source, multi-provider LLM agent runtime written in Rust, published under the MIT license. It runs an iterative agent loop — calling an LLM, executing tools on its behalf, feeding results back, and repeating until a task is complete. The project is maintained by weirdstuff.dev and is available on crates.io as well as via Docker or source build.

What It Is

Openheim is a general-purpose LLM agent that can be deployed in four distinct modes from a single binary: an interactive terminal UI (TUI), a headless CLI for one-shot prompts, an ACP stdio agent for editor integrations, or a self-hosted ACP-over-WebSocket server. It is provider-agnostic, supporting OpenAI, Anthropic Claude, Google Gemini, Ollama, vLLM, LM Studio, and any OpenAI-compatible endpoint. The agent loop, tool execution, conversation memory, and MCP integration are all shared across every run mode — no code changes needed when switching deployment targets.

Architecture and Run Modes

The single openheim binary exposes four entry points:

• Interactive TUI — a ratatui-based terminal UI with theming, session restore (:sessions), and live context-size tracking in the footer
• Headless CLI (openheim run) — streams a single prompt to stdout; supports model overrides and skill flags
• ACP stdio agent (openheim acp) — implements the Agent Client Protocol over stdio for native editor integration with Zed, JetBrains, and any ACP-compatible client
• Self-hosted server (openheim serve) — exposes ACP over a multiplexed WebSocket plus a REST API on port 1217; multiple clients connect simultaneously with isolated session contexts

The Rust implementation is highlighted for low memory footprint, fast startup, true concurrency, and memory safety by default.

Built-in Tools and MCP Extension

Openheim ships eight built-in tools available to the LLM: execute_command, read_file, write_file, edit_file, list_dir, search (ripgrep-style), web_fetch, and delegate_task for subagent delegation. Shell execution is disabled by default and must be explicitly enabled via allow_shell = true. File access is sandboxed to a configurable work_dir.

External tools are added by connecting MCP servers in config.toml — both stdio and Streamable HTTP transports are supported. Tools are auto-discovered at startup and namespaced as server__tool, requiring no code changes.

Skills System and Conversation Memory

The skills system lets users drop Markdown files into ~/.openheim/skills/ and reference them by name with --skills. They are injected into the system prompt at runtime after the agent's base identity (system.md). Default skills can be set in config.toml to auto-load every session; ACP clients can also pass skills per-session via the _meta field.

Every conversation — including full tool call history — is persisted to disk as JSON in ~/.openheim/history/. Sessions can be resumed from the TUI, headless CLI, or over ACP with session/load, continuing exactly where they left off.

Deployment Options

Openheim supports four deployment paths:

• Local binary — cargo install openheim from crates.io, or build from source
• Docker — multi-stage Dockerfile with a Compose file; mount a volume for history persistence
• Cloud / self-hosted — deploy the binary or Docker image on any Linux host (Fly.io, Railway, EC2, Raspberry Pi) behind a reverse proxy
• Rust library — embed the full agent runtime in any Rust application via the openheim crate with default-features = false to exclude TUI and server dependencies

Update: v0.9.0

The latest release is v0.9.0, published on September 1, 2026. The repository was created in January 2026 and has seen active development, with the most recent push on the same date as the v0.9.0 release. The crates.io page lists the current library version as 0.6 for embedding, while the binary release tracks the v0.9.x series. The project direction signals continued investment in ACP transport, MCP integration, and the Rust library embedding path.