EX MCP Server

# EXAI‑MCP Chat Sheet (As‑Configured) This cheat sheet captures how your EXAI‑MCP server will behave given your current .env and provides a fast operator reference plus a routing decision tree. ## Quick Start and Smoke - Start server: `python -m server` - Verify: call `status`, `version`, `listmodels` - Smoke: `python tools/ws_daemon_smoke.py` (checks streaming, web‑cue, long‑context) - Tail logs (PowerShell): - `Get-Content -Path logs\mcp_activity.log -Tail 200 -Wait` - `Get-Content -Path logs\mcp_server.log -Tail 200 -Wait` ## As‑Configured Behavior (from .env) - Agentic routing: `DEFAULT_MODEL=auto`, `ROUTER_ENABLED=true`, `EX_ROUTING_PROFILE=balanced` - Providers: KIMI and GLM enabled (keys present; allowlists set for safety) - Long‑context bias: `WORKFLOWS_PREFER_KIMI=true`, `EX_PREFER_LONG_CONTEXT=true` - Web search defaults ON: `EX_WEBSEARCH_ENABLED=true`, `EX_WEBSEARCH_DEFAULT_ON=true` (GLM/Kimi web browsing enabled) - Lean tool surface: `LEAN_MODE=true`, `STRICT_LEAN=true` (core tools only) - Expert pass off by default: `EXPERT_ANALYSIS_MODE=disabled`, `DEFAULT_USE_ASSISTANT_MODEL=false` - WS daemon reachable: `EXAI_WS_HOST=127.0.0.1`, `EXAI_WS_PORT=8765` - Structured logs + activity summaries: `LOG_FORMAT=json`, `EX_TOOLCALL_LOG_PATH=.logs/toolcalls.jsonl`, `EX_ACTIVITY_SUMMARY_AT_TOP=true` Effect: - Fast/default path = GLM `glm-4.5-flash` - Web/time cues keep GLM preferred for browsing - Long prompts bias Kimi/Moonshot; see thresholds below ## Routing Expectations - Web/time cues (URL, "today", or `use_websearch=true`): prefer GLM browsing path - Long‑context thresholds: - `estimated_tokens > 48k` → prefer Kimi/Moonshot - `estimated_tokens > 128k` → strongly prefer Kimi/Moonshot - Vision/multimodal: prefer GLM (configurable) - Default otherwise: GLM `glm-4.5-flash` Note: Upstream prompt shaping may shorten effective text. To steer reliably, pass explicit `estimated_tokens` (the client wrapper now does this automatically). ## Hints and Field Defaults (Client Wrapper Enhancements) - `estimated_tokens`: automatically injected from your prompt/messages before dispatch - `continuation_id`: automatically tracked and propagated across multi‑step workflow calls - Still recommended to set: - `use_websearch=true` when you want the browsing path - Include URLs or time‑sensitive wording for web cues ## Minimal Tool Examples (JSON‑style arguments) - chat: `{ "prompt": "Summarize README", "model": "auto" }` - thinkdeep (1 step): `{ "step": "Plan", "step_number": 1, "total_steps": 1, "next_step_required": false, "findings": "Kickoff" }` - codereview: `{ "step": "Review", "step_number": 1, "total_steps": 1, "next_step_required": false, "findings": "Start", "relevant_files": ["/abs/path/file.py"] }` ## Logs to Watch - `MCP_CALL_SUMMARY`: shows final `model`, `tokens`, `duration` - Quick filters: - `Select-String -Path logs\mcp_activity.log -Pattern "MCP_CALL_SUMMARY"` - `Select-String -Path logs\mcp_server.log -Pattern "ERROR|Traceback"` ## Troubleshooting Quick Picks 1) Route surprises → Check `MCP_CALL_SUMMARY`; ensure `estimated_tokens` is large enough; include more raw text if needed 2) No tools → Confirm `LEAN_MODE/LEAN_TOOLS/STRICT_LEAN` 3) Streaming errors → Use `stream_demo` (async streaming path implemented) 4) Multi‑step validation → Always include step metadata (step, step_number, total_steps, next_step_required, findings) ## Decision Tree (Routing) ```mermaid flowchart TD A[Tool call (model=auto)] --> B{Web/time cue? (use_websearch=true, URL, 'today')} B -- Yes --> G[Route to GLM browsing path] B -- No --> C{estimated_tokens > 128k?} C -- Yes --> H[Strongly prefer Kimi/Moonshot] C -- No --> D{estimated_tokens > 48k?} D -- Yes --> I[Prefer Kimi/Moonshot] D -- No --> E{Vision signal?} E -- Yes --> J[Prefer GLM (vision path)] E -- No --> F[Default GLM (glm-4.5-flash)] %% Notes subgraph Notes N1[estimated_tokens now sent by client wrapper] N2[LEAN_MODE=true restricts visible tools only] N3[Expert pass off by default; enable per-call if needed] end ``` ## Acceptance Checks (Day‑1/Day‑2) - stream_demo: fallback (zai) OK; streaming (moonshot) returns first chunk - thinkdeep with a URL or `use_websearch=true`: completes and routes per rules - chat_longcontext: executes; route depends on `estimated_tokens` (now injected) ## Safety & Defaults - Secrets only in `.env` (never commit); `.env.example` documents placeholders - Safe‑by‑default verification runs: linters/tests are OK without extra permission - Expert second‑pass is OFF unless you request it per‑call --- Maintained by EXAI‑MCP. Update this sheet when routing thresholds or provider preferences change.