PoopRusteek — Providers & MCP

LLM providers

The built-in DeepSeek web client is the default backend (cookie/token + local PoW, no API key). When you need more, /providers manages any number of extra endpoints speaking three wire protocols:

• OpenAI Chat Completions — LM Studio, Ollama /v1, vLLM, OpenRouter, …
• Anthropic Messages — the Claude API and compatible proxies
• Google Generative Language — Gemini

Adding & switching

/providers                 # open the panel: Enter activates, a adds, d removes
/providers add             # step-by-step wizard
/providers add <name> [openai|anthropic] <base_url> [model] [api_key]   # one-liner
/models                    # pick a model from the active provider's live list
/models <model_id>         # switch directly (validated against the live list)

Notes:

• The wizard's Model step is optional. An entry with an empty model serves the fetched list only; a bare <name> routes to the first fetched model, and activating that entry auto-opens the /models picker.
• Extra endpoints persist in config.toml as [[providers]] entries (protocol = "openai" | "anthropic" | "gemini"); active_provider picks the live one.

Model-list freshness

Provider model lists are fetched and cached in data_dir/provider_models.json.

/refetch-providers <ms|off>   # background refetch period (default 180000; off = startup + on-add only)
/cache-providers  <ms|off>    # how long the persisted cache stays valid across restarts (default 180000)

Bare /refetch-providers (or /cache-providers) shows both knobs plus per-provider fetch state (count, age).

MCP tool servers

MCP lets external servers expose tools to the agent. Tool names are namespaced mcp__{server}__{tool}. Browse connected servers and their tools with /mcp.

Discovery — 8 config sources (first-found-wins)

1. Own — {data}/pooprusteek/mcp.json (also stores enabled/disabled state)
2. Workspace — ./mcp.config.json, ./.vscode/mcp.json
3. Global — {config}/pooprusteek/mcp.config.json
4. Claude Desktop — claude_desktop_config.json
5. VS Code — settings.json → mcp.servers
6. Claude CLI — ~/.claude/mcp.json
7. Cursor — ~/.cursor/mcp.json
8. Opencode — {config}/opencode/…

Adding a server (/mcp add)

Three entry points, all converging on the same connect-and-persist path (written to mcp.json as an "Own" server):

/mcp add                                # 2-way modal: paste JSON | wizard
/mcp add <name> <command> [args…]       # one-line stdio shorthand (npx handled on Windows)
/mcp add {"mcpServers": {"name": {…}}}  # paste a Claude-Desktop-style JSON blob

The quick shorthand only covers stdio servers. For http/sse transports, or servers needing env vars / headers, use the wizard or paste JSON (any parse failure falls back to the wizard modal with an explanation).

Transports: stdio (subprocess, JSON over stdin/stdout, 60s/req), http (POST JSON, tracks MCP-Session-Id, 30s), sse (text/event-stream).

Cache & reload

/mcp ttl <secs>   # tools/list cache TTL (1–86400; default 300), persisted to config
/mcp reload       # force a fresh tools/list on all servers

OAuth (/mcp auth)

For HTTP/SSE servers that answer 401, /mcp auth (alias /mcp oauth) filters the list to servers needing authorization; Enter on one runs the full RFC 9728/8414/7591

• PKCE flow — it opens your browser, captures the redirect on a local port, and exchanges tokens. Tokens are stored encrypted in the OS keyring (never in mcp.json), and any client rebuild (connect/toggle/reconnect) picks them up automatically, refreshing when near expiry. Only authorization servers supporting RFC 7591 dynamic client registration work today; one requiring a pre-registered client_id fails with a clear error. Linux needs a working Secret Service backend.

Semantic tip

With many tools connected, PoopRusteek defers MCP schemas — the system prompt carries only a per-server summary (playwright (25 tools)), and full definitions arrive via per-turn hints or the tool_search builtin. See the pooprusteek-skills-and-rag skill.