More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Atomic Decomposition Proxy

    1

    This proxy fixes that by **forcing a three-phase process** so each step is simple enough to be solved well — without switching to an expensive model.

    $11.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Atomic Decomposition Proxy

    Atomic Decomposition Proxy

    Example session with this skill installed

    A client sends a compound request: 'Write a Python function that validates an email address, then write a unit test for it, then explain the regex in one sentence.' The proxy decomposes it into three atomic tasks, executes each in its own call, and synthesizes one final answer — instead of letting the small model try to do all three at once and fail.

    • Read your context and instructions
    • Compiled the atomic decomposition proxy

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

    About this skill

    Small and cheap models fail on compound tasks because they try to solve everything in one shot: they mix independent objectives, hallucinate data they don't have, or produce a half-finished answer. This proxy fixes that by forcing a three-phase process so each step is simple enough to be solved well — without switching to an expensive model.

    It is a drop-in replacement for any OpenAI-compatible endpoint. Point your client at the proxy instead of the upstream and nothing else changes.

    Every request passes through exactly three phases, orquestated by AtomicDecompositionEngine:

    1. Decompose — the model decides whether the instruction is atomic (solvable in one step) or should be split into concrete, ordered subtasks. It recurses up to a max depth, building a task tree.
    2. Execute atomic leaves — each leaf task is solved in its own model call, with accumulated context from earlier results. A leaf may invoke real tools across several rounds. If it needs an external action and has no matching tool, it emits a [[NECESITA_HERRAMIENTA: ...]] marker instead of hallucinating.
    3. Synthesize — all leaf results are consolidated into one final answer, presented as if solved directly (no mention of the internal decomposition). It can resolve any pending [[NECESITA_HERRAMIENTA]] marks with available tools.

    It is a proxy, not a model: it adds latency (one extra round-trip per phase) and token cost. For trivial single-step requests the overhead may not pay off — use it for genuinely compound tasks. Decomposition quality depends on the upstream model's judgment; the depth cap and tool policy bound the damage. Session state is in-memory only: restart the proxy and all sessions are lost (by design, no persistence). The proxy trusts the caller's system prompt as authority.

    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 any payment.

    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

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 15 days ago

    Requires

    Deepseek
    Python 3.10+
    FastAPI
    Uvicorn
    HTTPX
    Pydantic
    pydantic-settings
    OpenAI-compatible endpoint
    Python

    Permissions

    Allowed Hosts

    api.deepseek.com

    File Scopes

    reads
    its
    own
    .env
    file
    (if
    present)
    and
    bundled
    prompt
    .md
    files
    writes
    nothing
    to
    disk
    —
    all
    session
    state
    is
    in-memory

    Purpose: an OpenAI-compatible proxy that decomposes compound tasks into atomic ones. Network: makes HTTPS requests ONLY to the user-configured upstream endpoint (UPSTREAM_BASE_URL, default api.deepseek.com) via httpx; no other hosts, no scraping, no third-party APIs. File I/O: reads .env and prompt files only; never writes. Env vars: UPSTREAM_API_KEY (required), plus optional UPSTREAM_BASE_URL, UPSTREAM_MODEL, PROXY_HOST/PORT, MAX_DECOMPOSITION_DEPTH, MAX_TOOL_ROUNDS_PER_PHASE, REQUEST_TIMEOUT_SECONDS, SESSION_TTL_SECONDS, MAX_SESSIONS, EXPOSE_REASONING_CONTENT. No eval(), no shell subprocesses, no curl|sh, no rm -rf, no base64 blobs.

    Listed15 days ago
    Updated7 days ago

    What's inside

    Frequently Asked Questions

    No. It stretches ONE model's capability by decomposing compound tasks into atomic ones. It does not spawn or coordinate separate agents — it is a single-model proxy.

    Cost and speed. A cheap/small model that fails on multi-step work, run through this proxy, often matches a bigger model's output for a fraction of the cost — because each atomic step is simple enough to be solved well.

    The leaf task invokes the matching tool across several rounds. If it needs an external action with no matching tool, it emits a [[NECESITA_HERRAMIENTA: ...]] marker instead of hallucinating. The proxy pauses (finish_reason: tool_calls), your client runs the tool and sends the result back, and the proxy resumes exactly where it left off.

    No — session state is in-memory only with a TTL. Restarting the proxy loses all sessions by design (no disk writes). This keeps it simple and side-effect-free.

    No. The proxy prepends the caller's real system prompt as an authority layer over each internal phase prompt, so an agent client keeps its own conventions.

    New in LLM Ops