MCP server for Agence — AI governance framework with policy enforcement, skills, memory, and multi-LLM consensus
Critical failures across naming, descriptions, and schemas. 29 tools with minimal or missing documentation. Most tool names lack action verbs (launch_tool, mode_system, mode_knowledge are vague). Descriptions exist but are often generic or incomplete (e.g., 'System command dispatcher' without explaining WHEN to use it). Input schemas present for most tools but lack proper type constraints and validation guidance. No output schemas documented. Error handling absent. This server prioritizes internal shell-script composition over LLM-friendly tool design.
Appends decision records to the append-only Merkle-chained audit ledger with operation type and context
Displays current Agence state including root, config, LLM provider, and operational status
Hands off current session state to another agent with context preservation
Indexes multiple sources in batch with formtag classification and checksum generation
Creates indexed entry with checksum and semantic formtag hierarchy for knowledge sources
Data-driven tool launcher that reads binary, flags, and model configuration from registry.json and executes external tools (aider, claude, pilot) with proper environment setup and model override support
Returns availability status for all supported LLM providers with detection of API keys and active provider
Vague mode_* tool names lack action verbs. 'mode_system', 'mode_knowledge', 'mode_session', 'mode_external', 'mode_init' do not convey what action occurs. LLMs cannot infer intent from 'mode_' prefix alone.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., agent_id after handoff_to_agent). Responses are unstructured.
Descriptions are generic and lack WHEN/WHY guidance. 'System command dispatcher' (mode_system), 'Session lifecycle management' (mode_session) do not explain when to call vs alternatives or what prerequisites exist.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
External slash-command pass-through and git shortcut dispatcher for routing commands to external tools
Agence control-plane initialization and configuration dispatcher for ^init, ^reload, ^install, ^recon, ^scan commands
Knowledge management dispatcher for task/workflow/project/lesson/log/plan/todo/note/fault/issue/job matrix operations
Session lifecycle management including save, commit, push, handoff, pickup, pause, and resume operations
Displays command prefix symbol table and mode documentation
System command dispatcher that routes tool launches, persona agents, ensembles, and system utilities based on registry-driven type detection with unified route parsing for model qualifiers
Parses named model tier aliases (opus, sonnet, haiku) to full model names
Parses unified route syntax with dot-separated qualifiers and model tier resolution (EBNF: target ('.' qualifier)*)
Pauses current session and saves state for later resumption
Picks up a handoff from another agent and resumes session with transferred context
Reindexes knowledge bases with optional source type and output format specification
Resolves the appropriate LLM model based on provider, operational mode (query/plan/code), and blast radius with explicit override support
Detects and returns the active LLM provider based on environment variables, config file, or API key presence with explicit override support
Comprehensive route resolution combining provider detection, model selection, mode, blast radius, and cost tier calculation
Resumes a previously paused session with full context restoration
Sends a query to the configured LLM provider and prints the response to stdout with multi-provider support (Anthropic, OpenAI, Azure, Gemini, Mistral, Groq, OpenRouter, Grok, Qwen, Copilot, Cline, Ollama)
Shows provider availability table with detection status for all supported LLM providers
Detects and exports LLM provider and model configuration from environment variables, config file, or auto-detection based on available API keys
Sends a planning request to the LLM and returns structured ACTION/CONFIDENCE/STEPS response block
Indexes and scans knowledge bases, codebase files, and world model entities with semantic formtag classification
Displays comprehensive help text for Agence command modes, tool launchers, and system utilities
Displays Agence version information
Parameter descriptions lack format/constraint guidance. 'command' params in mode_session, mode_init, mode_external accept free-form strings with no enum or validation hints. LLMs will hallucinate invalid commands.
No error handling or recovery guidance. Tools do not document what errors are retryable, user-fixable, or fatal. No actionable error messages defined.