Agentlas-OS

Agentlas OS is a Python engine (the Hephaestus plugin) that staffs a team of specialist agents onto a job from inside a coding-agent host. It runs as a plugin or command adapter for Claude Code, Codex, Gemini CLI, Antigravity, Cursor, OpenCode, OpenClaw, Hermes, and compatible local or API hosts. A companion desktop app, Agentlas Desktop, offers the same engine behind a window on macOS, Windows, and Linux.

Work moves through three steps. A sentence is split into tasks with roles, a handoff order, and the artifact each task owes, and an org chart whose review points back to an earlier task is rejected before anything runs. Candidate agents for each task then come back from the local machine, the owner's private Agent Cloud, and the public Agentlas Hub at once, in an unranked list: the host model picks, and its reason is recorded. Qualification has four levels (declared, checked, demonstrated, attested), each task can set a floor, and the chosen release is pinned by digest so later drift stops the run. Planning, work, synthesis, and verification run as separate invocations, an independent verifier has to pass the result, and each handoff is proven by receipt or reported with the state it stopped in (prepared, blocked, source_unavailable).

Agents and the protocol for calling them follow the A2A v1.0 standard. Agent cards are read and written at the standard well-known path, an external agent is callable only when its capability aligns and its identity is verified by a signed card, imports act as proposals, exports redact private paths and logic, and invocations are caller-gated. Candidate rows are compressed (one task with 20 candidates drops from 40,873 to 10,087 bytes), and the project reports no ranking by design: the schema pins history evidence to false and a source returning a scored list has its answer refused.

An agent package is a method document, not a worker: the LLM is supplied by the user (bring your own model), the runtime facilities (web, files, shell, browser) by the host machine, and the package carries the procedure, judgment rules, source priority, input and output contracts, and stop conditions. A meta-agent factory compiles single agents, multi-agent teams (PM orchestrator, memory curator, policy gate, QA), workspace bundles, and agents built from an exported working session, with raw transcripts, hidden prompts, credentials, and private paths excluded from the package. Agent Cloud stores and retrieves the owner's packages but never executes them: the selected model and current host run the work under that host's permission model.

The desktop app brings every model the user already has into one place: the coding CLIs they subscribe to (Claude Code, Codex, Antigravity and Gemini, Cursor, Grok, Kimi), their own API keys across 11 providers, and local models through a bundled llama.cpp engine that pulls GGUF models from Hugging Face, plus LM Studio and MLX. When the active model runs out of quota, the turn is re-run on the next model in the user's priority order and the original returns when the limit clears. Further subsystems include Stormbreaker (execution gating with verified, unverified, and blocked completion states and a resumable run journal), an ontology runtime that indexes local files into an FTS5 and SQLite store with reciprocal rank fusion and GraphRAG search, and governed memory that enforces agent scope, privacy, expiry, and curator approval before ranking.