jev-delegate
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@jev-delegate Rank the search results for 'memory leak' by relevance to our issue."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Jev Delegate for Codex and Claude Code
Jev Delegate gives Codex and Claude Code two local MCP tools: search_and_rank ranks large rg result sets, and classify_items assigns caller-defined labels to English text. Local search collects candidates; TypeSafe Jev through OpenRouter judges them. It cannot browse, write code, summarize, or act as a general subagent.
Each installation uses its own OPENROUTER_API_KEY. No key belongs in this repository or an agent prompt. OpenRouter may charge for requests.
Install for Codex
Requires Node.js 20+, npm, rg, and Codex CLI. Get your own OpenRouter key from OpenRouter.
git clone https://github.com/koosbcom/jev-delegate.git
cd jev-delegate
npm ci
npm run check
codex mcp add jev-delegate -- node "$PWD/dist/server.js"
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
cp -R skills/jev-delegate "${CODEX_HOME:-$HOME/.codex}/skills/"Set the key in the environment that starts Codex. For a Codex CLI session, enter it without putting its value in shell history:
read -s OPENROUTER_API_KEY
export OPENROUTER_API_KEY
codexFor Codex desktop on macOS, set it for the app session, then restart Codex:
read -s OPENROUTER_API_KEY
launchctl setenv OPENROUTER_API_KEY "$OPENROUTER_API_KEY"
unset OPENROUTER_API_KEYStart a new Codex task after installation. If your MCP client does not expose workspace roots, set JEV_DELEGATE_WORKSPACE_ROOT to the project directory in the Codex process environment. Check registration with codex mcp list.
Related MCP server: openrouter-jev-mcp
Install for Claude Code
Requires Node.js 20+, npm, rg, and Claude Code. Build once:
git clone https://github.com/koosbcom/jev-delegate.git
cd jev-delegate
npm ci
npm run checkRegister the MCP server and the skill for every project:
claude mcp add --scope user jev_delegate -- node "$PWD/dist/server.js"
mkdir -p ~/.claude/skills
cp -R skills/jev-delegate ~/.claude/skills/Or load the repository as a plugin for one session. .claude-plugin/plugin.json bundles the skill and the MCP server:
claude --plugin-dir /path/to/jev-delegateExport OPENROUTER_API_KEY in the shell that starts claude, as shown above. Claude Code exposes the project directory as the MCP workspace root, so JEV_DELEGATE_WORKSPACE_ROOT is not needed. Check registration with claude mcp list or /mcp. Telemetry goes to the same ~/.codex/state/jev-delegate/usage.jsonl path unless JEV_DELEGATE_TELEMETRY_PATH is set.
Opening this repository itself in Claude Code also offers the project .mcp.json server; it fails to start until npm run build has created dist/.
When to use
search_and_rank: more than 20 workspace search hits or roughly 4,000 raw-result tokens. Supply explicit literal or regex patterns and a relevance criterion.classify_items: many English items with 2–20 distinct labels and clear label criteria.Keep small, sensitive, non-English, and open-ended work with the main model.
The server returns confidence, cost metadata, and a bounded raw fallback if Jev fails. Treat decisions as evidence, not proof. The search tool reads only the current workspace, respects ignore rules, excludes common credential paths, and screens snippets for secret-like text. These filters are heuristic; review material before sending it to any external API.
Default spend guards are $0.01 per tool call and $0.25 per day. The server writes local metadata-only telemetry to ~/.codex/state/jev-delegate/usage.jsonl; set JEV_DELEGATE_TELEMETRY=off to disable it. See src/config.ts for optional limits.
Development
npm ci
npm run checkOffline tests and an MCP stdio test run in npm run check. An opt-in live test needs your own key: npm run test:live. The live API path has not yet been verified against a user's OpenRouter key, so API behavior may need adjustment.
License
MIT. See LICENSE.
Available Tools
2 toolsclassify_itemsClassify supplied items with JevARead-onlyIdempotent
Assign exactly one caller-supplied label to each item using Jev typed decisions. Adds other and insufficient_context labels. Use for English bulk classification, not prose generation.
Inputs: items (required), labels (required), instructions (required). Returns: usage, status, classifications, fallback_reason
search_and_rankSearch workspace and rank matches with JevARead-onlyIdempotent
Run read-only ripgrep inside the current workspace, then use Jev to return only relevant matches. Use automatically when raw search would exceed 20 candidates or about 4K tokens. English judgments only.
Inputs: patterns (required), relevance_criterion (required), mode, top_k, +2 more. Returns: usage, status, results, truncated, +3 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 2 tools
The two tools are completely distinct: one focuses on searching and ranking relevant matches from raw search results, while the other assigns labels to items. There is no overlap in purpose, and an agent can easily select the correct tool based on the task.
Both tools use an imperative verb phrase, but 'search_and_rank' is a compound verb while 'classify_items' follows a verb_noun pattern. The slight inconsistency is minor and does not hinder readability, but a more uniform pattern (e.g., 'search_and_rank_items') would be ideal.
With only two tools, the server feels thin for a delegation service. However, the narrow scope of search ranking and classification justifies a small count, making it borderline rather than excessive. Each tool serves a clear purpose, so the count is acceptable but leaves little room for additional functionality.
The tool surface covers the core functions implied by the server name: search-and-rank and classification. Minor gaps exist (e.g., no summarization or generation tool), but the stated purpose is well-addressed, and an agent can complete common tasks without hitting dead ends.
Maintenance
Related MCP Connectors
Jev-powered decisions, web search, PDF/web to Markdown, summarize. From $0.001, no API key.
Grade essays, check resumes, and ask typed yes/no, choice and rubric questions, powered by Jev.
Typed-question decisions from Jev: classify, moderate, route models and guard tool calls.
BYOK AI router & FinOps control plane: audit token waste, check tool safety, and rank models.
Related MCP Servers
- AlicenseAqualityCmaintenanceEnables Codex to rank available tools and skills, rerank code and documentation search results, and triage saved output artifacts using Jev-based relevance scoring while leaving final decisions to Codex.17MIT
- AlicenseBqualityCmaintenanceEnables coding agents to obtain probabilistic decisions from Jev AI via OpenRouter for classification, scoring, and validation, with tools like jev_check, jev_classify, jev_score, and jev_evaluate.53MIT
- AlicenseAqualityBmaintenanceConnect AI agents to Jev AI (jev-ai.pro) for classification, scoring, yes/no checks, advisory action assessment, batched typed decisions and saved judges. Returns structured answers and probabilities using your Jev AI API key.6MIT
- AlicenseAqualityBmaintenanceEnables coding agents to query files, logs, and search results through TypeSafe Jev's typed, probabilistic answers, so they retrieve only the needed conclusion instead of raw context. Supports classification, scoring, yes/no checks, extraction, page ranking, and injection screening.11209 npmMIT