chore: clean up MCP config, add mcp-server research notes

Author: localpibox

.pi/mcp.json

@@ -1,7 +1,3 @@
 {
-  "mcpServers": {
-    "exa": {
-      "disabled": true
-    }
-  }
+  "mcpServers": {}
 }

GH-PROFILE-DRAFT.md

@@ -0,0 +1,205 @@
+# MCP Server Research Summary
+
+> Researched: 2026-08-07
+
+---
+
+## 1. context7 (upstash/context7)
+
+| Field | Value |
+|---|---|
+| **URL** | `https://github.com/upstash/context7` |
+| **Stars** | 60,396 ⭐ |
+| **Forks** | 2,901 |
+| **Language** | TypeScript |
+| **License** | MIT |
+| **Created** | 2025-03-26 |
+| **Last Pushed** | 2026-08-07 (today) |
+| **Open Issues** | 31 |
+
+### Maturity: 🟢 HIGH
+- Massive adoption (60k+ stars) — the most popular MCP server by far
+- **101 releases**, MCP v2 support already shipped
+- Multi-language SDKs: TypeScript, Python (ai-sdk tools), Go (HTTP SSE/Streamable)
+- Multi-client support: Claude Code, Cursor, Copilot, VS Code, Codex, Gemini
+- **Packages:** `@upstash/context7-mcp`, `ctx7` CLI, SDK, AI SDK tools, **`@upstash/context7-pi`** (dedicated pi.dev extension)
+- Actively maintained by Upstash (well-funded company)
+- Weekly releases, daily commits
+- Clean issue tracker: mostly library report requests (doc content fixes)
+
+### Usefulness: 🟢 HIGH
+- **Solves:** LLM hallucination from outdated/wrong documentation by providing real-time API docs to AI agents
+- Directly addresses a core pain point: LLMs citing deprecated APIs or wrong function signatures
+- Integrates with Cursor, Claude Code, Windsurf, and any MCP client
+
+### Community Feedback: 🟢 POSITIVE
+- The 60k star count reflects widespread adoption
+- Community describes it as "the BEST MCP server for AI coding assistants"
+- Often recommended as a "power combo" with sequential-thinking
+- **Concern:** API key required; some users seek self-hosted alternatives
+  - Open-source alternative: `arabold/docs-mcp-server` (1,620 stars)
+
+### Verdict: **STRONG YES** — This is a must-have for any LLM dev stack. The primary alternative is "context7-like" but nothing else has comparable coverage.
+
+---
+
+## 2. github-mcp-server (github/github-mcp-server)
+
+| Field | Value |
+|---|---|
+| **URL** | `https://github.com/github/github-mcp-server` |
+| **Stars** | 32,034 ⭐ |
+| **Forks** | 4,748 |
+| **Official** | Yes (GitHub's own product) |
+| **Last Release** | v1.8.0 (2026-07-30) |
+| **Release Cadence** | ~weekly |
+| **Open Issues** | 362 |
+
+### Maturity: 🟢 HIGH
+- Official GitHub product — production-grade
+- Very active maintenance: weekly releases
+- 32k+ stars, strong community
+- **23+ toolsets:** repos, issues, PRs, actions, code_security, dependabot, discussions, gists, git, projects, notifications, orgs, users, labels, copilot, copilot_spaces, and more
+- **Toolset-level filtering** to reduce LLM context size
+
+### Usefulness: 🟢 HIGH
+- Full GitHub API access: repos, issues, PRs, branches, commits, actions, Copilot reviews
+- Directly supports dev workflows: create PRs, manage issues, review code, run Actions
+- Replaces manual GitHub CLI + manual git operations
+
+### Community Feedback: 🟡 MIXED (but positive overall)
+- **Pros:** Feature-complete, official support, well-maintained
+- **Cons:** Some bugs around cross-fork PR reviews, HTTP mode auth gaps, install docs issues
+- 362 open issues is a concern but most are feature requests, not blockers
+
+### Verdict: **YES** — If you do GitHub workflows, this is essential. However, it's a **large** server with many capabilities. If you only need basic repo/file access, it may be overkill. Your current stack doesn't use it yet.
+
+---
+
+## 3. fetch-mcp (zcaceres/fetch-mcp)
+
+| Field | Value |
+|---|---|
+| **URL** | `https://github.com/zcaceres/fetch-mcp` |
+| **Stars** | 809 |
+| **Forks** | 119 |
+| **Language** | TypeScript |
+| **Last Pushed** | 2026-03-12 (5 months ago) |
+| **Open Issues** | 5 |
+
+### Maturity: 🟡 MEDIUM
+- Moderate adoption (809 stars)
+- Last push was **March 2026** — development appears stalled
+- Recent security fixes (SSRF bypass, DNS rebinding) show awareness of risks
+- Small, focused codebase (~1MB)
+
+### Usefulness: 🟡 LOW-MEDIUM (for your stack)
+- **Solves:** Simple HTTP fetching of web content
+- Your stack already has **Exa MCP** (`web_fetch_exa`) which does this **much better** — it returns clean markdown, handles paywalls, and provides structured results
+- No clear differentiation over Exa for your use case
+
+### Community Feedback: 🟡 NEUTRAL
+- Small but dedicated user base
+- Issues are mostly security hardening and minor fixes
+- The existence of `fetcher-mcp` (1,070 stars, Playwright-based) as an alternative suggests the space is crowded
+
+### Verdict: **NO** — Redundant with Exa MCP. Your `web_fetch_exa` is superior for research use cases. Only consider if you need raw HTTP (not just content fetching) with custom headers/cookies.
+
+---
+
+## 4. filesystem MCP (official MCP reference implementation)
+
+| Field | Value |
+|---|---|
+| **URL** | Part of `github.com/modelcontextprotocol/servers` (`src/filesystem/`) |
+| **Parent Stars** | **89,300** ⭐ |
+| **npm** | `@modelcontextprotocol/server-filesystem` |
+| **Language** | TypeScript |
+| **License** | MIT |
+| **Official** | ✅ Anthropic/MCP team |
+| **Tools** | 13 specialized ops |
+
+### Maturity: 🟢 HIGH
+- Part of the **official MCP servers repo** — Anthropic/MCP team maintained
+- 89.3k stars on parent repo — highest adoption signal in the ecosystem
+- Published as npm package
+- 13 specialized tools with dry-run previews and safety annotations
+
+### Usefulness: 🟠 LOW (for your stack)
+- **13 tools:** read_text_file, write_file, edit_file (dry-run), search_files, directory_tree, and more
+- **Roots-based access control** — dynamically scoped directories for security
+- **MCP ToolAnnotations:** readOnlyHint, idempotentHint, destructiveHint — LLM-aware safety
+- **Your stack already has this built-in:** Pi.dev's native `read`, `bash`, `grep`, `find`, `write`, `edit` cover 100% of filesystem needs
+- An MCP filesystem would be **pure duplication**. The safety annotations are elegant but not worth the integration cost when Pi already handles it.
+
+### Community Feedback: 🟠 MIXED
+- The official version is well-designed but niche adoption as a standalone server
+- Community prefers alternative implementations (Rust version, workspace servers)
+
+### Verdict: **NO** — Pi's built-in tools (read, write, edit, bash, grep, find) are superior. The official MCP version's safety annotations (readOnlyHint, destructiveHint) are nice but not worth the integration overhead.
+
+---
+
+## 5. sequential-thinking (arben-adm/mcp-sequential-thinking)
+
+| Field | Value |
+|---|
+| **URL** | `https://github.com/arben-adm/mcp-sequential-thinking` |
+| **Stars** | 941 |
+| **Forks** | 115 |
+| **Language** | Python |
+| **Last Pushed** | 2026-08-07 (today) |
+| **Open Issues** | 1 |
+| **Description** | *(empty)* |
+
+### Maturity: 🟡 MEDIUM
+- Moderate adoption (941 stars)
+- Recently pushed but activity is **all dependency bumps** (chore-only)
+- No description on GitHub — suggests low effort
+- Forks exist with better descriptions (spences10/mcp-sequentialthinking-tools, 584 stars)
+
+### Usefulness: 🟡 LOW-MEDIUM
+- **Solves:** Structured, step-by-step reasoning for complex tasks
+- Your stack already has **thinking layers** (`reasoning_effort` parameter for Qwen models) which provides structured reasoning natively
+- An MCP server for "thinking" is meta: you're asking an AI to think about thinking
+- Useful only if you need to **serialize/inspect** the thinking process (rare)
+
+### Community Feedback: 🟡 NEUTRAL
+- Most commits are automated dependency updates
+- Multiple forks suggest the original is not the definitive version
+- No major complaints, but no enthusiastic endorsements either
+
+### Verdict: **MAYBE** — Only useful if you need an *inspectable, serializable* reasoning chain (e.g., for debugging agent decisions). For normal dev work, Qwen's native `reasoning_effort` is sufficient and simpler.
+
+---
+
+## Summary Matrix
+
+| MCP Server | Stars | Maturity | Usefulness | Recommendation | Why |
+|---|---|---|---|---|---|
+| **context7** | 60k | High | High | ✅ **STRONG YES** | Prevents LLM doc hallucination; unique value |
+| **github-mcp-server** | 32k | High | High | ✅ **YES** | Official GitHub integration; essential for GH workflows |
+| **fetch-mcp** | 809 | Medium | Low | ❌ **NO** | Redundant with Exa `web_fetch_exa` |
+| **filesystem** | 675 | Low | Low | ❌ **NO** | Pi has read/write/edit/grep/find built-in |
+| **sequential-thinking** | 941 | Medium | Low | ⚠️ **MAYBE** | Redundant with Qwen native reasoning; only if you need inspectable chains |
+
+### Ecosystem Maturity (2025-2026)
+
+The MCP ecosystem shows strong maturity:
+- **Official MCP servers repo:** 11.4k forks, 908 contributors, 4,158 commits
+- **Registry:** Central MCP Registry for curated server catalog
+- **Protocol:** MCP v2 announced and shipped
+- **Multi-language SDKs:** TypeScript, Python, Go, Java, Rust, C#, PHP, Ruby, Swift
+- **Vendor adoption:** GitHub, Cursor, Claude Code, Copilot, VS Code, Codex all support MCP
+- **Trend:** Specialized servers migrating to standalone repos (GitHub MCP now at `github/github-mcp-server`)
+
+### Notable Alternatives
+- **Self-hosted docs:** `arabold/docs-mcp-server` (1,620 stars) — open-source Context7 alternative
+- **Web search:** `exa` (already in your stack) beats Brave Search MCP for research
+- **Web fetch:** `fetch-mcp` (809 stars) is official reference but your Exa MCP is superior
+
+### Top Recommendations for Your Stack
+
+1. **context7** — Install immediately. This is the highest-ROI addition.
+2. **github-mcp-server** — Install if you do GitHub workflows (PRs, issues, etc.)
+3. Skip fetch-mcp, filesystem, and sequential-thinking — your stack already covers or exceeds these capabilities natively or via Exa MCP.