MCP orchestrator — connect unlimited Model Context Protocol servers through 8 meta-tools. Progressive tool discovery keeps your AI agent's context window small.
This MCP server presents a meta-orchestration tool with significant definition quality gaps. While tool names are generally action-oriented (search_tools, execute_tool, manage_mcps, auth_mcp), the descriptions range from adequate to verbose, and critical schema information is either missing or inferred from test files rather than visible in actual tool registration code. The server implements 10 tools with varying levels of documentation. Most critically, no actual tool registration code is visible in the provided source, only test files reference tool definitions, making it impossible to verify schemas are properly declared. Parameter descriptions exist but lack specificity around constraints, ranges, and valid values. Error handling descriptions are minimal. The run_slop tool description is exceptionally verbose (500+ chars) with embedded examples that violate the no-examples-in-descriptions principle.
Build shell command streaming agnt daemon events via `agnt monitor`. Returns `command` for Claude Code Monitor tool or any shell runner. Filters: target (errors/interactions/process/all), proxy_id, process_id, severity, format.
OAuth for MCP servers. Actions: login (start flow), logout (drop token), status (check), list (all authenticated). Returns text.
Override tool descriptions, define custom tools. Actions: set_override, remove_override, list_overrides, define_custom, remove_custom, list_custom, export, import.
Execute tool on MCP server. Passes parameters through. Returns underlying MCP response as-is.
Metadata for connected MCPs. Compact by default (names+descriptions). verbose=true for full schemas. mcp_name+tool_name for single-tool schema.
Manage MCP connections. Actions: register, unregister, reconnect, list, status, health_check, list_stale_overrides. Returns text.
No visible tool registration code in provided source. Tool definitions are inferred from test files (execute_tool_wrongkey_test.go), not from actual registration. Cannot verify schemas are properly declared in the MCP server implementation.
run_slop description is 500+ characters with embedded code examples ('data = api.fetch(id: 42)', 'results = []'), violating the principle that examples in descriptions cause LLMs to reuse them literally. Should be 50-150 chars with examples moved to separate documentation.
Parameter constraints are missing specificity. 'limit' parameters lack min/max (e.g., search_tools says 'default 20, max 100' but no minimum stated). 'action' enum parameters list valid values in description text rather than as formal JSON Schema enum constraints.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 56 | <=2025-11-25 | v2 |
| 2026-03-09 | C | 64 | - | v1 |
Execute SLOP script with access to all registered MCPs. Inline script or file path. Returns final expression value as text. Call MCP tools as mcp_name.tool_name(param: value). Example patterns: Chain results between tools: data = api.fetch(id: 42) summary = ai.summarize(text: data["content"]) emit(summary) Loop and collect: results = [] for id in [1, 2, 3]: results = results + [api.get(id: id)] emit(items: results, count: len(results)) Transform with builtins: repos = github.search(query: "mcp") names = map(repos, |r| r["name"]) emit(join(names, "\n")) Pipe for chaining transforms (left value becomes first arg): [1, 2, 3, 4, 5] | filter(|x| x > 2) | map(|x| x * 10) data | json_stringify() Session memory persists across run_slop calls (thread-safe): store_set("key", value) prev = store_get("key") Persistent memory survives restarts (disk-backed): mem_save("bank", "key", value, description: "what this stores") data = mem_load("bank", "key") entries = mem_list("bank") matches = mem_search("query") Use recipe parameter: recipe: "list" to see available templates, recipe: "<name>" to load one. Use slop_reference to browse built-in functions (map, filter, reduce, json_parse, regex_match, etc.).
Fuzzy search tools across connected MCPs. Ranked results. Paginated (default: 20, max: 100). Use offset for next page. Response includes total and has_more.
Full details for SLOP function by name. Returns formatted text.
Search SLOP built-in functions. Compact output (name+signature) by default. verbose=true for full details. list_categories=true for category counts.
execute_tool and manage_mcps accept 'mcp_name' and 'tool_name' parameters with no guidance on how to discover valid values. LLM must call search_tools or get_metadata first, but this dependency is not documented in parameter descriptions.
Error handling is generic. Tools return 'text' responses but provide no guidance on error classification (retryable vs user-fixable vs fatal). No error messages shown. LLM has no recovery path if a tool call fails.
manage_mcps 'register' action requires both 'command' and 'args' parameters but their relationship is undocumented. Are both required together? Mutually exclusive? What format does 'args' expect?
customize_tools combines 7 distinct operations (set_override, remove_override, list_overrides, define_custom, remove_custom, list_custom, export/import) into a single 'action' parameter. This violates single-responsibility principle, should be 7 separate tools or at minimum clearly documented sub-tool semantics.
agnt_watch is a discovery/read tool but its 'format' parameter is undocumented, what formats are valid? Unclear if this is JSON, plaintext, CSV, or something else.