Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Research & Engineering Router

    1

    Free MIT-licensed evidence screening for coding and research. Let your configured low-cost model select relevant sources before the main model reads them, while your main agent stays in charge.

    Free

    Security scanned

    Research & Engineering Router

    Example session with this skill installed

    Using the published paired benchmark evidence for jev-research-engineering, report the change in primary-model token usage for the code audit and research tasks, plus the change in whole-process time. State the scope and limitations.

    • Read your context and instructions
    • Compiled the research & engineering

    Recorded benchmark report

    This example report summarizes the repository's recorded paired benchmark; it is not a newly executed agent session.

    TaskPrimary model alonePrimary model with routingPrimary-model token reduction
    Code audit100,50686,48013.96%
    Research99,53193,1546.41%
    Combined200,037179,63410.20%

    The counted primary-model tokens are input, including cached input, plus output. Reductions are calculated as (baseline minus assisted) / baseline. Research alone did not meet a 10% reduction target.

    Mean end-to-end time was 66.45 seconds for the main model alone and 80.02 seconds with routing: 20.42% longer for the complete workflow, including the helper.

    The evaluation used two authored tasks and three paired repetitions per task. All six assisted runs passed the authored acceptance checks. The coding task evaluated an audit, not a complete implementation. These results do not establish the same benefit on arbitrary repositories or research questions, and do not imply a lower fixed subscription fee.

    Next step: run a paired evaluation on your own task before relying on the reported benefit. Preserve decisive dependencies and contrary evidence, and have the lead model verify the final result. Full setup, audit evidence and methodology are available at https://github.com/MehradYaghoubi/jev-research-engineering.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Keep the main model focused

    Research & Engineering Router is one free, MIT-licensed skill (installation slug: jev-research-engineering) for coding and research. A configured inexpensive model screens large evidence sets before the main model reads them. Your main model retains reasoning, implementation, conclusions and verification; there is no independent supervisor. Jev is optional.

    What it helps with

    • Code audits and development: relevant source selection, Clean Code, meaningful OOP and checks for missing route, startup and event connections.
    • Research: source-based evidence selection, provenance and retention of contrary evidence.
    • Auditing: record context selection, provider usage and fallback behavior.

    Run screening before broad evidence intake. Small, obvious tasks should skip routing. The skill cannot remove context the main model has already read, and selected evidence still needs verification.

    Measured results

    These are measured primary-model input tokens (including cached input) plus output tokens, not a promise for every task.

    Recorded taskMain model aloneWith routingReduction
    Code audit100,50686,48013.96%
    Research99,53193,1546.41%
    Combined200,037179,63410.20%

    Mean whole-process time increased from 66.45 to 80.02 seconds: 20.42% longer. Evaluation: two authored tasks, three paired repetitions each; all six assisted runs passed the authored quality checks. The coding case was a code audit, not a full autonomous implementation. Results depend on task, model, cache and provider. A fixed ChatGPT/Codex subscription bill does not fall automatically when token usage falls.

    Install: choose local OR global

    Project-local:

    typed command: npx skills add MehradYaghoubi/jev-research-engineering --skill jev-research-engineering

    Global: use the same command with --global. Choose your agent in the installer. Avoid installing both copies for the same host.

    Codex invocation: $jev-research-engineering Claude Code standalone invocation: /jev-research-engineering Pi invocation: /skill:jev-research-engineering

    Connect your model

    Python 3.10+ and permission to run the included helpers are required. Configure a provider endpoint, model and your own credential using the setup guide. Supported interfaces are native System One, OpenAI-compatible Chat Completions JSON and Anthropic Messages. Examples cover GapGPT, OpenRouter, Anthropic-compatible services and local Ollama. GapGPT has live verification; other configurations have offline coverage and are not all live-certified.

    The skill is free. Your chosen provider may charge separately. Credentials are never included in this download. An Agensi/MCP host must expose the helper files and allow Python execution; otherwise use the direct GitHub installation.

    Documentation and video

    GitHub: installation, measured results and silent video guide

    Provider setup

    Installation video with captions

    Thank you for trying the skill. If it helps your work, consider starring the original GitHub repository or sharing feedback. A star is never required.

    Runtime permissions and lightweight package

    The Agensi download contains 28 text/source files with standard-library Python helpers, setup references, examples and the measured report. Development tests and the binary installation video stay on GitHub. SKILL.md declares exactly what is read/written, the selected credential variable, optional private Windows DPAPI setup, and the configured inference endpoint. Only approved task evidence is sent to that provider; no analytics or unrelated credential collection is included. These permissions are necessary for API-assisted screening and remain visible in the listing.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Benefits depend on task size, cache, model and provider. Small tasks should skip routing. Context already read by the main model cannot be removed. Screening can omit important evidence, so the lead agent must verify dependencies and contrary sources and fall back when needed. The published benchmark covers two authored tasks with three paired repetitions, not every coding/research workflow. Research alone showed 6.41% reduction. Whole-process mean time increased 20.42%. Requires Python execution and helper-file access; not every agent/provider combination has been live-tested. Provider API charges are separate from this free MIT skill.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free to use

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 19 days ago

    • Free with an Agensi account

    Software
    Python 3.10+Terminal access
    System One
    orOpenAI
    orAnthropic API
    Git
    Optional
    Node.js

    Any OpenAI-compatible endpoint works. Add your own provider key where needed (config, environment or Windows DPAPI). Node.js is only for the optional Skills CLI install.

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables

    Required: run the four bundled standard-library Python helpers; read only evidence approved for this task; write new audit files and explicitly selected private configuration outside the skill folder. Screening sends that task/evidence only to the endpoint in the selected provider profile (default example: https://api.typesafe.ai/v1/systemone; other profiles are optional). Only the profile-selected credential variable is used, or optional Windows DPAPI; no collection of other provider keys. DECISION_GATE_CONFIG/JEV_GATE_CONFIG optionally select the private profile. HTTPS/authentication required remotely; explicit literal loopback HTTP supported locally. No telemetry, dependency installation, shell evaluation, MCP server or lifecycle hook is bundled. GitHub links/video are documentation, not runtime API destinations. Node.js/Git are optional installation tools, not Python-runtime requirements.

    Listed19 days ago

    What's inside

    Frequently Asked Questions

    Yes. The full skill and source are free under the MIT license. Jev is optional: configure a compatible model/provider or supported local endpoint. Provider usage may incur separate charges; no credential or paid API credit is bundled.

    Compatible SKILL.md hosts with Python 3.10+, file access and permission to run included helpers, including documented Codex, Claude Code and Pi installation paths. Supported interfaces include native System One, OpenAI-compatible Chat Completions JSON and Anthropic Messages. Only GapGPT has live verification so far; not every host/provider combination is certified.

    The Agensi ZIP includes 28 text/source files: SKILL.md, four standard-library Python helpers, provider/setup references, example profiles and evidence, the historical measured report, MIT license, privacy guidance and agent metadata. Full development tests and the silent captioned installation video are available on the linked GitHub repository rather than bundled in this lightweight ZIP. The main agent remains responsible for conclusions and changes.

    In two authored benchmark tasks with three paired repetitions each, primary-model input including cached input plus output fell 13.96% for code audit and 6.41% for research; combined reduction was 10.20%. Mean complete-process time rose 20.42%. These are recorded results, not guaranteed savings or a reduction in fixed subscription fees.

    Use npx skills add MehradYaghoubi/jev-research-engineering --skill jev-research-engineering for project-local installation, or add --global for global installation. Choose one scope for your host. Follow the linked GitHub setup guide to configure endpoint, model and credential privately. Invoke the installed skill using the command documented for your agent.

    Yes. The workflow asks the lead agent to preserve mandatory sources and contrary evidence, expand missing dependencies, verify selected evidence and fall back when routing fails. It cannot guarantee that every necessary source was selected, so verification remains essential.

    New in LLM Ops

    Free