S18Share is an open-source AI agent runtime and orchestration API. Use POST /runs to execute workflows with MCP tools, memory, and RAG. Quick start: uv sync, set GEMINI_API_KEY, run uv run python api.py, then open /docs. For full architecture, integration, deployment, and operations details, see docs/ and docs/QUICKSTART_5_MIN.md.
Query routing, skill matching, and planner fast-paths can use TypeSafe AI's Jev ("System One") model for typed decisions with calibrated confidence — falling back to the built-in regex guards whenever Jev is unkeyed, unreachable, or low-confidence. Set JEV_API_KEY (or TYPESAFE_API_KEY) to enable; S18_DECISIONS_ENABLED=0 disables it at runtime. Tune the decisions block in config/settings.json, and validate your key with python scripts/jev_probe.py "state".
Planner agents can drive the user's real, logged-in Chromium through mcp_servers/server_browserskill.py, a safety proxy over Tencent's BrowserSkill bsk CLI. Work runs in a separate Agent Window using the browser's existing session — no credentials are extracted — and the user's own tabs must be explicitly borrowed. Tools follow the CLI (bsk_session_start, bsk_navigate, bsk_observe, bsk_click, bsk_fill, bsk_screenshot, bsk_tab_borrow, bsk_request_help, and the rest) behind a default denylist, a JSONL audit log at data/system/browserskill_audit.jsonl that redacts form values, and output paths confined to data/system/browserskill/. Install the CLI and browser extension first (Windows: irm https://raw.githubusercontent.com/Tencent/BrowserSkill/main/install.ps1 | iex); without them the server still starts and bsk_cli_status reports the capability as unavailable. This complements the anonymous server_browser.py — use web_search/web_extract_text for public research and bsk_* only when a task needs the user's login. Tuning knobs are documented in mcp_servers/README_browserskill.md.
Planner agents can drive the desktop on Windows/macOS/Linux through mcp_servers/server_computer.py, a safety proxy over the Cua Driver. It discovers the driver's tools at runtime and re-exposes them as computer_<name> (computer_click, computer_type_text, and so on) behind a default denylist, an optional allowlist, and a JSONL audit log at data/system/computer_audit.jsonl. Install the driver first (Windows: irm https://cua.ai/driver/install.ps1 | iex); without it the server still starts and exposes only computer_driver_status and computer_refresh_tools. Computer use is a vision task, so route these runs to the Gemini arm or another strong vision model rather than the small Ollama default. Tuning knobs (S18_COMPUTER_TOOL_DENYLIST, S18_COMPUTER_PERMISSION_MODE, and the rest) are documented in mcp_servers/README_computer.md.