Open-source multi-client AI coding harness and terminal agent with shared rules, local/hosted model routing, memory, MCP and safety policy.
Agnostic AI combines two related layers. The harness compiles a shared rules, traits and safety-policy source into 18 AI coding clients, including Claude Code, Codex, Cursor, Windsurf, Copilot, Cline, Aider, OpenHands and others. Its terminal agent adds file, shell and search tools, subagents, code indexing, checkpoints, testing workflows, persistent memory, MCP servers and model switching across local and hosted endpoints. 1
The current repository documents Python 3.9+ and Node 18+ requirements, Windows development plus Ubuntu CI, and local dashboards bound to loopback. These are documented capabilities rather than AiToolMap hands-on findings. 1
The software is free and MIT-licensed. Local-model operation can use LM Studio or Ollama; configured hosted models, provider APIs or external subscriptions can create separate third-party costs. AiToolMap found no separate paid Agnostic AI software plan on the verified repository surface. 1
The project documents a shared fail-closed policy layer, secret-path blocking, human confirmation for hard-stop commands unless explicitly trusted, backups for rules synchronization, and local-only dashboard binding with mutation-token/origin protections. These design controls are useful and unusually explicit for a small open-source project, but they are not an independent security assessment. 1
The canonical 50-source panel found no exact independent review, benchmark or eligible numeric rating corpus for Agnostic AI. Limited community discussion exists outside the fixed panel, but it is too small and informal to convert into numeric evidence. No hands-on testing was performed. 2
Agnostic AI is most relevant to developers already juggling several coding agents and configuration formats. The shared-rules harness directly targets configuration drift, while the terminal agent is useful for teams or individuals that want to move between local and hosted models without changing the surrounding workflow. 1
It is less suitable for buyers who need enterprise support, a large user community, independently established reliability or a mature hosted product with published SLAs.
MIT source; broad 18-client rules synchronization; local and hosted model routing; substantial terminal-agent tooling; MCP and persistent memory; explicit guard/policy architecture; local dashboards; active repository history. 1
small public adoption footprint; no exact independent benchmark/review corpus found; limited external evidence for long-run reliability and safety; operational breadth creates configuration complexity and powerful shell/file capabilities still require careful workspace and permission discipline. 1
Products matched by shared tasks, audience, workflow, product type and capabilities.