Zephex

Zephex MCP

Ten tools at https://zephex.dev/mcp for the user's project — not training data.

They may only say "use Zephex" or "use MCP." You choose the tool and call it (find_code, read_code, …). Some editors show zephex:find_code — same tool.

Workflow

| Situation | Chain | |-----------|--------| | New repo | get_project_context → optional explain_architecture | | Feature, fix, refactor | scope_task → read_code on focus_files | | Unknown location | find_code → read_code | | Large rename | find_code + exhaustive: true → files_summary | | Add or upgrade dependency | check_package → audit_package if needed | | Stuck after 2+ tries | keep_thinking ↔ find_code / read_code | | Production URL | audit_headers | | Generic patterns (not their repo) | Zephex_dev_info search → get |

path on every repo tool: absolute folder (/Users/jane/api) or github:owner/repo. Monorepo: path = app root, not empty parent.

inline_files: only without disk access — { "rel/path": "" }.

Max 3 calls of the same tool per turn without new evidence.

Step 1 — Orient or scope

New repo or vague ask:

get_project_context({ "path": "/abs/project", "detail_level": "brief" })

Before reading many files for a task:

scope_task({
  "task": "Add rate limiting to POST /api/upload",
  "path": "/abs/project",
  "hint_symbols": ["handleUpload"]
})

Use focus_files, callers_at_risk, suggested_approach. Good task: one clear sentence. Bad: "fix it".

Step 2 — Read or search

Known symbol:

read_code({ "target": "handleUpload", "path": "/abs/project" })

Unknown location:

find_code({ "query": "validateToken", "path": "/abs/project", "scope": "definitions" })

Then read_code on the best hit. Skip find_code if you already know the symbol.

Step 3 — Package, architecture, URL, reasoning

check_package({ "package": "express", "ecosystem": "npm" })

explain_architecture({ "path": "/abs/project", "focus": "auth" })

audit_headers({ "url": "https://example.com" })

keep_thinking({
  "thought": "Hypothesis: rate limit runs after handler",
  "thoughtNumber": 1,
  "totalThoughts": 5,
  "nextThoughtNeeded": true,
  "confidence": 0.6,
  "thoughtType": "hypothesis",
  "goalAnchor": "Add rate limiting to POST /api/upload",
  "lastActions": ["find_code(query=rateLimit)"]
})

Tool name is keep_thinking, not thinking.

Zephex_dev_info({ "operation": "search", "query": "Stripe webhook signature" })

Then operation: "get" with the slug from search.

Tool reference

| Tool | When | Key args | |------|------|----------| | get_project_context | New session | path, force? | | scope_task | Task before bulk reads | task, path, hint_symbols?, max_files? (1–15) | | find_code | Don't know where | query, path, scope?, exhaustive? | | read_code | Known symbol/files | target or mode+files, path | | check_package | Before install | package, ecosystem | | audit_package | Upgrade/CVE | package, task, from_version? | | explain_architecture | How this repo works | path, focus? | | audit_headers | User gave URL | url | | keep_thinking | Stuck / risky change | thought, thoughtNumber, goalAnchor, lastActions | | Zephex_dev_info | Playbooks | operation: search → get |

read_code callers / blast_radius: local path only — on github: use find_code usages.

If they only said "use Zephex"

1. Guess the goal from their message.
2. Pick a Workflow row.
3. Call with path + specific args.
4. Answer from tool JSON only.

Errors

| Error | Action | |-------|--------| | Unauthorized | npx zephex setup | | 429 | Tell user; retry once | | Missing path | Ask absolute project directory |

Mistakes

• Install without check_package
• Many reads without scope_task
• find_code when target is known
• thinking instead of keep_thinking
• Repo answers from memory when Zephex is connected

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zephexMCP
• Source: zephexMCP/agent-skills
• License: MIT
• Homepage: https://zephex.dev

Install and usage instructions live in the source repository linked above.