Open-source configuration-first agent layer for local LLM runtimes with tools, isolated memory and deny-first permissions over OpenAI-compatible endpoints.
LocalHarness is an MIT-licensed, model-agnostic agent layer for locally served LLMs. It sits above OpenAI-compatible runtimes such as vLLM, Ollama, LM Studio and llama.cpp and defines agents in YAML rather than requiring users to write a Python framework. The current repository documents hierarchical agents, built-in file/shell/Python/web tools, MCP, per-agent and per-project SQLite memory, append-only event logs, workspace layering, subagents and optional Discord control. 1
The project also implements tool-calling fallback for models without native function calling, endpoint/model discovery, context/token accounting, a scenario benchmark suite and an autoresearch loop for gated harness experiments. These are documented project capabilities, not AiToolMap hands-on findings. 1
LocalHarness is free open-source software under the MIT License. The software itself has no verified paid tier; users supply their own local or self-hosted inference runtime and therefore bear the hardware, electricity or hosting costs associated with that setup. 1
The current repository requires Python 3.12+ and uv. Linux is exercised by CI; Windows is supported natively with Git for Windows; the harness can connect over HTTP to a model server on another machine. 1
LocalHarness documents a deny-first permission architecture in which inherited policies can only narrow, workspace-level deny patterns union rather than widen, and untrusted web content is structurally separated from host-mutating tools. It also includes a SECURITY.md and explicitly warns users to secure remotely reachable inference endpoints rather than exposing unauthenticated servers. 1
The repository reports hermetic tests, opt-in live-runtime tests and maintainer validation across local runtimes and reference hardware. These are meaningful engineering signals, but AiToolMap treats them as first-party evidence until independently reproduced. The project itself labels the current v0.13.2 release early-stage/pre-1.0. 1
The canonical 50-source panel found no exact independent review, benchmark or eligible numeric rating corpus. No AiToolMap hands-on testing was performed.
LocalHarness fits developers and technically capable teams that already operate local inference and want a config-first agent layer with tools, memory, permissions and orchestration. It is especially relevant for recurring or unattended jobs where local compute, data locality and freedom from token metering are important. 1
It is less suitable for users who want a hosted turnkey assistant, enterprise SLA/support commitments, or independently established benchmark leadership.
explicit MIT license; broad local-runtime compatibility; YAML configuration; hierarchical agents; MCP; per-project memory isolation; deny-first permission model; extensive repository history; tests and benchmark infrastructure; Windows/Linux/macOS support; clear networking/security caveats. 1
pre-1.0 interfaces may change; no exact independent panel evidence found; performance/runtime-validation claims are maintainer-generated; self-hosting creates hardware and operations burden; local agents with shell/file tools still require disciplined permission and endpoint security. 1
Products matched by shared tasks, audience, workflow, product type and capabilities.