Narsil MCP - Deep Code Intelligence

Rust-powered code intelligence providing 76 specialized tools for security scanning, call graph analysis, and structural queries. Accessed via Code Mode for token-efficient on-demand access (~700 tokens vs ~6,000-8,000 native).

Core Principle: Unified code intelligence - Narsil handles STRUCTURE, SECURITY, and SEMANTIC search.

---

1. 🎯 WHEN TO USE

Activation Triggers

Use when:

• Security vulnerability scanning needed (OWASP, CWE, injections)
• Call graph or control flow analysis required
• Finding symbols, definitions, or references
• Dead code detection or complexity analysis
• SBOM generation or license compliance
• Git blame, history, or hotspot analysis

Keyword Triggers:

• Security: "security scan", "vulnerability", "injection", "OWASP", "CWE", "taint"
• Call Graph: "call graph", "callers", "callees", "who calls", "what calls"
• Structure: "find symbols", "find functions", "project structure", "references"
• Quality: "dead code", "unreachable", "complexity", "unused"
• Supply Chain: "SBOM", "dependencies", "licenses", "CVE"
• Git: "git blame", "hotspots", "contributors", "history"

Use Cases

Security Auditing

• Full security scan with OWASP Top 10 and CWE Top 25
• SQL injection, XSS, command injection detection
• Taint analysis tracing untrusted data flow
• Secret and credential detection

Code Understanding

• Project structure overview
• Symbol search (functions, classes, structs)
• Call graph visualization
• Dependency analysis

Code Quality

• Dead code and unreachable code detection
• Complexity metrics (cyclomatic, cognitive)
• Unused assignments and uninitialized variables

Supply Chain Security

• SBOM generation (CycloneDX, SPDX)
• CVE checking against OSV database
• License compliance verification

When NOT to Use

Do not use for:

• Simple text pattern search → Use Grep
• File path/name matching → Use Glob

For semantic search: Use narsil_neural_search - Narsil's neural backend handles meaning-based queries.

---

2. 🧭 SMART ROUTING

Activation Detection

TASK CONTEXT
    │
    ├─► Security scanning needed
    │   └─► Load: references/security_guide.md
    │       └─► Tools: scan_security, find_injection_vulnerabilities, trace_taint
    │
    ├─► Call graph / code flow analysis
    │   └─► Load: references/call_graph_guide.md
    │       └─► Tools: get_call_graph, get_callers, get_callees
    │
    ├─► Quick overview / first use
    │   └─► Load: references/quick_start.md
    │       └─► Tools: get_project_structure, find_symbols
    │
    ├─► Full tool reference needed
    │   └─► Load: references/tool_reference.md
    │       └─► All 76 tools documented
    │
    └─► Simple structural query
        └─► Use SKILL.md only
            └─► Tools: find_symbols, get_project_structure

Resource Router

def route_narsil_resources(task):
    """
    Resource Router for mcp-narsil skill
    Load references based on task context
    """

    # ──────────────────────────────────────────────────────────────────
    # SECURITY SCANNING
    # Purpose: OWASP, CWE, injection detection, taint analysis
    # Key Insight: Start with scan_security for overview, drill into specifics
    # ──────────────────────────────────────────────────────────────────
    if task.involves_security or task.mentions_vulnerability:
        return load("references/security_guide.md")  # Security workflow

    # ──────────────────────────────────────────────────────────────────
    # CALL GRAPH ANALYSIS
    # Purpose: CFG, DFG, callers/callees, complexity metrics
    # Key Insight: Start with get_call_graph, then drill with callers/callees
    # ──────────────────────────────────────────────────────────────────
    if task.involves_call_graph or task.mentions_code_flow:
        return load("references/call_graph_guide.md")  # Analysis workflow

    # ──────────────────────────────────────────────────────────────────
    # QUICK START
    # Purpose: First-time usage, verification, basic commands
    # Key Insight: Fastest path to working state
    # ──────────────────────────────────────────────────────────────────
    if task.is_first_use or task.needs_verification:
        return load("references/quick_start.md")  # Getting started

    # ──────────────────────────────────────────────────────────────────
    # COMPLETE REFERENCE
    # Purpose: All 76 tools with descriptions and priority
    # Key Insight: Use for discovery or when unsure which tool
    # ──────────────────────────────────────────────────────────────────
    if task.needs_tool_discovery or task.needs_full_reference:
        return load("references/tool_reference.md")  # All tools

    # Default: SKILL.md covers basic structural queries

# ══════════════════════════════════════════════════════════════════════
# STATIC RESOURCES (always available, not conditionally loaded)
# ══════════════════════════════════════════════════════════════════════
# assets/tool_categories.md → Priority categorization of all 76 tools
# scripts/update-narsil.sh   → Script to update Narsil binary

---

3. 🛠️ HOW IT WORKS

Code Mode Invocation

Narsil is accessed via Code Mode's call_tool_chain() for token efficiency.

Naming Convention:

narsil.narsil_{tool_name}

Process Flow:

STEP 1: Discover Tools
       ├─ Use search_tools() for capability-based discovery
       ├─ Use tool_info() for specific tool details
       └─ Output: Tool name and parameters
       ↓
STEP 2: Execute via Code Mode
       ├─ Use call_tool_chain() with TypeScript code
       ├─ Await narsil.narsil_{tool_name}({params})
       └─ Output: Tool results
       ↓
STEP 3: Process Results
       └─ Parse and present findings

Tool Invocation Examples

Important: Most Narsil tools require a repo parameter. Use list_repos() to discover the repo name (typically "unknown" if not explicitly named).

// Discover security tools
search_tools({ task_description: "security vulnerability scanning" });

// Get tool details
tool_info({ tool_name: "narsil.narsil_scan_security" });

// First: Discover repo name
call_tool_chain({
  code: `
    const repos = await narsil.narsil_list_repos({});
    return repos;  // Returns repo name (e.g., "unknown")
  `
});

// Execute security scan (repo required)
call_tool_chain({
  code: `
    const findings = await narsil.narsil_scan_security({
      repo: "unknown"
    });
    return findings;
  `
});

// Find all functions (repo required, use symbol_type not kind)
call_tool_chain({
  code: `
    const symbols = await narsil.narsil_find_symbols({
      repo: "unknown",
      symbol_type: "function"
    });
    return symbols;
  `
});

// Get call graph (repo required, use function not function_name)
call_tool_chain({
  code: `
    const graph = await narsil.narsil_get_call_graph({
      repo: "unknown",
      function: "main"
    });
    return graph;
  `
});

// Get symbol definition (repo required, use symbol not name)
call_tool_chain({
  code: `
    const def = await narsil.narsil_get_symbol_definition({
      repo: "unknown",
      symbol: "myFunction"
    });
    return def;
  `
});

Unified Code Search

Narsil provides both structural AND semantic search capabilities:

Query Type	Tool	Reason
"How does authentication work?"	narsil_neural_search	Semantic meaning
"Find code similar to this"	narsil_neural_search	Vector similarity
"List all auth functions"	narsil_find_symbols	Structural query
"Scan for vulnerabilities"	narsil_scan_security	Security analysis
"Show call graph for login"	narsil_get_call_graph	Code flow

Neural Semantic Search

Narsil supports three neural embedding backends for semantic code search:

Backend	Model	API Key	Dimensions	Best For
Voyage AI	voyage-code-2	VOYAGE_API_KEY	1536	Code search (RECOMMENDED)
OpenAI	text-embedding-3-small	OPENAI_API_KEY	1536	General purpose
Local ONNX	Built-in	None required	384	Offline/privacy

Configuration Examples:

# Voyage AI (recommended for code)
--neural --neural-backend api --neural-model voyage-code-2
# Requires: VOYAGE_API_KEY in .env

# OpenAI
--neural --neural-backend api --neural-model text-embedding-3-small
# Requires: OPENAI_API_KEY in .env

# Local ONNX (no API key needed)
--neural --neural-backend onnx
# No API key required - runs locally

Tool: narsil_neural_search Fallback: BM25 search if neural unavailable

// Example: Semantic code search
call_tool_chain({
  code: `
    const results = await narsil.narsil_neural_search({ 
      query: "how does authentication work",
      top_k: 10 
    });
    return results;
  `
});

See references/tool_reference.md for complete tool documentation.

Error Handling Patterns

Narsil tools can fail for various reasons. Use these patterns for robust error handling.

Pattern 1: Check Index Status Before Search

call_tool_chain({
  code: `
    const status = await narsil.narsil_get_index_status({});
    
    if (!status.neural_ready || status.embedding_count === 0) {
      console.log("Neural index building, using structural query");
      const symbols = await narsil.narsil_find_symbols({
        repo: "unknown",
        symbol_type: "function",
        name_pattern: "auth"
      });
      return { fallback: true, results: symbols };
    }
    
    const results = await narsil.narsil_neural_search({
      query: "authentication flow"
    });
    return { fallback: false, results };
  `
});

Pattern 2: Try-Catch with Retry

call_tool_chain({
  code: `
    try {
      const findings = await narsil.narsil_scan_security({ repo: "unknown" });
      return { success: true, findings };
    } catch (error) {
      if (error.message?.includes("not found")) {
        await narsil.narsil_reindex({});
        const findings = await narsil.narsil_scan_security({ repo: "unknown" });
        return { success: true, retried: true, findings };
      }
      throw error;
    }
  `,
  timeout: 60000
});

Pattern 3: Graceful Degradation

call_tool_chain({
  code: `
    async function robustSearch(query) {
      try {
        const neural = await narsil.narsil_neural_search({ query, top_k: 10 });
        if (neural.results?.length > 0) return { method: "neural", results: neural.results };
      } catch (e) { console.log("Neural unavailable:", e.message); }
      
      try {
        const semantic = await narsil.narsil_semantic_search({ repo: "unknown", query });
        if (semantic.results?.length > 0) return { method: "semantic", results: semantic.results };
      } catch (e) { console.log("Semantic unavailable:", e.message); }
      
      const symbols = await narsil.narsil_find_symbols({
        repo: "unknown",
        name_pattern: query.split(" ")[0]
      });
      return { method: "symbols", results: symbols };
    }
    
    return await robustSearch("authentication handler");
  `
});

---

4. 📋 RULES

✅ ALWAYS

1. ALWAYS use Code Mode for Narsil invocation

   • Call via call_tool_chain() with TypeScript
   • Saves ~5,300 tokens vs native MCP

2. ALWAYS use full tool naming convention

   • Format: narsil.narsil_{tool_name}
   • Example: narsil.narsil_scan_security({})

3. ALWAYS use neural_search for semantic queries

   • "How does X work?" → narsil_neural_search
   • "Find code like X" → narsil_neural_search

4. ALWAYS load security_guide.md for security tasks

   • Provides phased workflow with checkpoints
   • Ensures comprehensive coverage

5. ALWAYS verify tool exists before calling

   • Use search_tools() or tool_info() first
   • Prevents "tool not found" errors

❌ NEVER

1. NEVER skip the narsil_ prefix in tool names

   • Wrong: await narsil.scan_security({})
   • Right: await narsil.narsil_scan_security({})

2. NEVER use Narsil's LSP/remote tools unnecessarily

   • LSP: IDE handles this natively
   • Remote: Not needed for local development

3. NEVER assume index is current

   • Use narsil_reindex({}) after file changes
   • Index persists with --persist flag

⚠️ ESCALATE IF

1. ESCALATE IF large repository causes timeout

   • Increase timeout: { timeout: 120000 }
   • Use --persist flag to save index
   • Consider indexing subset of files

2. ESCALATE IF tool returns unexpected results

   • Verify tool name with tool_info()
   • Check if repository is indexed with list_repos()
   • Trigger reindex() if needed

3. ESCALATE IF security findings seem incomplete

   • Verify all categories enabled in scan
   • Check language support (15 languages)
   • Consider taint analysis for deeper inspection

⚠️ KNOWN LIMITATIONS

1. JavaScript Call Graph

   • Call graph analysis may return empty or incomplete results for JavaScript
   • tree-sitter-javascript has limited support for dynamic call patterns
   • Workaround: Use find_symbols + get_symbol_definition for JS code analysis

2. Security Scanning Language Support

   • Security rules are primarily designed for backend languages (Rust, Python, Go, Java, C/C++)
   • JavaScript/TypeScript security coverage is limited
   • Frontend-specific vulnerabilities (DOM XSS) may not be fully detected

3. Neural Search After Index Clear

   • Clearing .narsil-index/ requires OpenCode restart to rebuild neural embeddings
   • Simply calling reindex() may not regenerate embeddings if MCP server has stale state
   • Fix: Exit OpenCode (Ctrl+C), then restart

4. Git Blame Unicode (Fixed in local build)

   • Original Narsil had issues with files containing Unicode box-drawing characters
   • Local build includes fix for char-safe string slicing

5. Code Mode Process Spawning

   • Code Mode spawns fresh Narsil MCP processes for each call batch
   • Search indexes (BM25, TF-IDF, Neural) are in-memory and rebuilt on each spawn
   • Impact: Semantic search may return empty if index hasn't built yet (~40-60s)
   • Workaround 1: Keep related calls in same call_tool_chain() batch
   • Workaround 2 (Recommended): Use HTTP server mode for reliable search (see below)

HTTP Server Workaround for Search

For reliable search functionality, use Narsil's HTTP server instead of Code Mode:

# Start HTTP server (indexes build once and stay warm)
./.opencode/skill/mcp-narsil/scripts/narsil-server.sh start

# Wait ~60s for indexes, then search reliably
./.opencode/skill/mcp-narsil/scripts/narsil-search.sh semantic "query"
./.opencode/skill/mcp-narsil/scripts/narsil-search.sh neural "how does X work"
./.opencode/skill/mcp-narsil/scripts/narsil-search.sh symbols function

# Or call any tool via HTTP
curl -X POST http://localhost:3001/tools/call \
  -H "Content-Type: application/json" \
  -d '{"tool": "semantic_search", "args": {"repo": "unknown", "query": "test"}}'

Scripts available:

• narsil-server.sh - Server management (start/stop/restart/status/logs)
• narsil-search.sh - Search CLI (neural/semantic/code/hybrid/symbols)

1. Functions with Limited Results

   • get_project_structure: May return empty tree structure
   • find_references: Returns 0 for most symbols (needs call graph traversal)
   • get_chunk_stats: Returns 0 (aggregate stats not populated)
   • hybrid_search: Requires BOTH BM25 + Neural indexes to be built

2. Code Mode Config Loading (CRITICAL)

   • Code Mode loads .utcp_config.json at startup only
   • Changes to Narsil config require OpenCode restart to take effect
   • If narsil is not defined error: restart OpenCode
   • Verify after restart: code_mode_list_tools() should show narsil.* tools
   • Common config issues: extra fields, relative paths, missing transport: "stdio"

Function Categories by Index Dependency

Category	Functions	Timing
AST-Based (work immediately)	find_symbols, get_symbol_definition, get_file, get_chunks, find_dead_code, scan_security, get_recent_changes	Instant
Index-Based (need ~40-60s)	semantic_search, search_code, hybrid_search, neural_search, find_similar_code	40-60s after server start

Tip: Use AST-based functions for immediate results. For semantic search, either wait ~60s after first Narsil call or trigger reindex() and wait.

---

5. 🏆 SUCCESS CRITERIA

Security Audit Complete

Security audit complete when:

• ✅ scan_security executed with OWASP/CWE rules
• ✅ find_injection_vulnerabilities checked SQL/XSS/command
• ✅ trace_taint analyzed untrusted data flow
• ✅ generate_sbom created dependency manifest
• ✅ check_dependencies verified against CVE database
• ✅ check_licenses confirmed compliance
• ✅ All critical/high findings addressed or documented

Code Analysis Complete

Code analysis complete when:

• ✅ get_project_structure provided overview
• ✅ find_symbols identified key components
• ✅ get_call_graph mapped function relationships
• ✅ find_dead_code identified cleanup candidates
• ✅ get_complexity flagged refactoring targets
• ✅ narsil_neural_search used for semantic understanding

Validation Checkpoints

Checkpoint	Validation
tools_discovered	search_tools() returns Narsil tools
repo_indexed	list_repos() shows repository
scan_complete	Security scan has zero unaddressed criticals
analysis_complete	All structural queries executed

---

6. 🔌 INTEGRATION POINTS

Framework Integration

This skill operates within the behavioral framework defined in AGENTS.md.

Key integrations:

• Gate 2: Skill routing via skill_advisor.py
• Tool Routing: Per AGENTS.md Section 6 decision tree
• Memory: Context preserved via Spec Kit Memory MCP

Code Mode Integration

Configuration: Add to .utcp_config.json:

CRITICAL:

• Use absolute path for command (e.g., /Users/username/bin/narsil-mcp)
• Do NOT include extra fields like _note, _neural_backends in mcpServers
• Config changes require OpenCode restart to take effect

{
  "name": "narsil",
  "call_template_type": "mcp",
  "config": {
    "mcpServers": {
      "narsil": {
        "transport": "stdio",
        "command": "/absolute/path/to/narsil-mcp",
        "args": [
          "--repos", ".",
          "--index-path", ".narsil-index",
          "--git",
          "--call-graph",
          "--persist",
          "--watch",
          "--neural",
          "--neural-backend", "api",
          "--neural-model", "voyage-code-2"
        ],
        "env": {
          "VOYAGE_API_KEY": "${VOYAGE_API_KEY}"
        }
      }
    }
  }
}

Finding your Narsil path:

which narsil-mcp
# Use the output as the absolute path in config

After config changes:

1. Save .utcp_config.json
2. Restart OpenCode (Ctrl+C, then restart)
3. Verify: code_mode_list_tools() should show narsil.* tools

Note: The --http flag enables a visualization web UI at localhost:3000, NOT HTTP transport for MCP. MCP transport is always stdio.

Index Persistence

Narsil supports persisting the code index to disk for faster startup:

Flag	Purpose	Default
--persist	Enable saving/loading index from disk	Disabled
--index-path	Custom index storage location	~/.cache/narsil-mcp/

Project-local indexes (recommended): Use --index-path .narsil-index and add .narsil-index/ to .gitignore.

Manual save: Use narsil.narsil_save_index({}) via Code Mode to trigger index save.

HTTP Server & Visualization

Narsil includes an HTTP server with React frontend for interactive graph visualization.

Important: The --http flag enables a visualization web UI, NOT HTTP transport for MCP. MCP communication is always via stdio - Code Mode spawns the Narsil process and communicates via stdin/stdout. The HTTP server runs in parallel for visual exploration.

Starting the HTTP Server (CRITICAL):

The HTTP server requires stdin to stay open. Use this pattern to keep it running:

# Backend (port 3000) - MUST use stdin pipe to prevent EOF shutdown
(tail -f /dev/null | narsil-mcp \
  --repos . \
  --index-path .narsil-index \
  --git --call-graph --persist \
  --http --http-port 3000 > /tmp/narsil-http.log 2>&1) &

# Frontend (port 5173) - in separate terminal
cd /path/to/narsil-mcp/frontend && npm install && npm run dev

Why the pipe? MCP servers read from stdin. Without input, they receive EOF and shut down. The tail -f /dev/null | keeps stdin open indefinitely.

Graph Views & Limitations:

View	Purpose	Required Parameters
import	Module import/export	None (best for JS)
call	Function call graph	None
symbol	Symbol definitions	Requires root parameter (function name)
hybrid	Combined import + call	Requires repo parameter
flow	Data flow	None

Known Frontend Issues:

• Symbol view shows "Symbol view requires a root" - enter a function name in the Root field
• Hybrid view may fail without repo parameter - use Import or Call view instead

Performance: Use Limit slider in UI, index specific dirs with -r src -r lib, add **/node_modules to .gitignore.

Related Skills

Skill	Integration
mcp-code-mode	Tool orchestration - Narsil accessed via Code Mode's call_tool_chain()
system-spec-kit	Context preservation - save security findings for future sessions

Tool Usage Guidelines

Bash: Execute update script, verify binary Read: Load reference files for detailed workflows

Binary Location

The Narsil MCP server source is embedded in the skill folder:

.opencode/skill/mcp-narsil/mcp_server/

Binary path (after build):

.opencode/skill/mcp-narsil/mcp_server/target/release/narsil-mcp

The absolute path is configured in .utcp_config.json. A symlink at ~/bin/narsil-mcp also points to this binary.

To update Narsil:

bash .opencode/skill/mcp-narsil/scripts/update-narsil.sh

---

7. 🏎️ QUICK REFERENCE

Essential Commands

Task	Tool	Example
Semantic search	neural_search	narsil.narsil_neural_search({ query: "authentication flow" })
Find symbols	find_symbols	narsil.narsil_find_symbols({ symbol_type: "function" })
Security scan	scan_security	narsil.narsil_scan_security({ repo: "." })
Call graph	get_call_graph	narsil.narsil_get_call_graph({ function: "main" })
Find callers	get_callers	narsil.narsil_get_callers({ function: "login" })
Find references	find_references	narsil.narsil_find_references({ symbol: "UserService" })

Common Patterns

// Semantic code search
call_tool_chain({
  code: `
    const results = await narsil.narsil_neural_search({
      query: "how does authentication work"
    });
    return results;
  `
});

// Security scan
call_tool_chain({
  code: `
    const findings = await narsil.narsil_scan_security({
      repo: "."
    });
    return findings;
  `
});

// Find all functions
call_tool_chain({
  code: `
    const functions = await narsil.narsil_find_symbols({
      symbol_type: "function"
    });
    return functions;
  `
});

Troubleshooting

Issue	Solution
Neural search returns empty	Wait 60s after server start for index to build
Call graph empty for JS	Known limitation - use find_symbols + find_references instead
Tool not found	Check naming: narsil.narsil_{tool}
Slow first query	Index building in progress - check with get_index_status

Error Quick Reference

Error Pattern	Cause	Solution
"not found" in message	Repo not indexed	Run narsil.narsil_reindex({})
"timeout" in message	Large codebase	Use categories filter, increase timeout
neural_ready: false	Index building	Wait ~60s or use AST-based tools
Empty results	No matches or index issue	Try different search method

---

8. 🔗 RELATED RESOURCES

scripts/

Script	Purpose	Usage
update-narsil.sh	Update to latest version	bash .opencode/skill/mcp-narsil/scripts/update-narsil.sh
narsil-server.sh	HTTP server management	bash scripts/narsil-server.sh start|stop|status
narsil-search.sh	Search CLI	bash scripts/narsil-search.sh neural "query"

references/

Document	Purpose	Key Insight
tool_reference.md	All 76 tools documented	Complete parameter reference
security_guide.md	Security scanning workflow	OWASP/CWE checkpoints
call_graph_guide.md	Call graph analysis	CFG/DFG patterns
quick_start.md	Getting started	5-minute setup

assets/

Asset	Purpose
tool_categories.md	Priority categorization of all 76 tools

External Resources

• Narsil GitHub - Source code and documentation
• OWASP Top 10 - Security standard
• CWE Top 25 - Weakness enumeration

Related Skills

• mcp-code-mode - Tool orchestration (Narsil accessed via Code Mode)
• system-spec-kit - Context preservation across sessions

Install Guide

• MCP - Narsil.md - Installation and configuration