Jungle Monkey

Agent orchestrator · Nº 01

Every coding agent. One control room.

Run Claude Code, Codex and Cursor side by side — queued, sandboxed and on a budget — on your server or your own machines. Watch every agent work, answer its questions, and ship the result as a branch or a pull request.

Claude Code · Codex · Cursor · parallel runs · sealed containers · self-hosted · MIT

#42 · Retry webhook deliveriesClaude Code

QUEUED

waiting for a slot

    Queued$0.000s
    #43 · Add CSV export to tablesCodex

    QUEUED

    waiting for a slot

      Queued$0.000s
      #44 · Upgrade ESLint to v9Cursor

      QUEUED

      waiting for a slot

        Queued$0.000s

        § 01 · Providers

        Any agent

        Three coding agents, one workflow. Each runs through its own SDK in the same sealed container, reports into the same timeline, gets the same skills and MCP tools, and delivers the same way.

        Claude Code

        Through the Claude Agent SDK, with an API key or your Claude subscription. Stops to ask when it needs a decision, and reports its own cost (API-equivalent on a subscription).

        Codex

        Through the OpenAI Codex SDK. Makes its own choices and lists open questions in the summary; its cost is estimated from the price table you can edit.

        Cursor

        Through the Cursor SDK. Like Codex, it decides on its own and lists open questions at the end, with an estimated cost.

        • Pick per project or per task. A project has a default agent and any task can override it, so one project can mix providers. Schedules pick theirs too.
        • Keys stay yours. Provider keys live on your own hosts: the server's .env, the OS keychain for npx jungle-monkey, the macOS Keychain on a computer you add. A run gets only its own agent's key.

        § 02 · Orchestration

        Orchestration

        A queue in front of every agent, limits on what runs at once and what it may spend, and loops that bring a task home.

        Queue and concurrency

        Tasks queue and start when their host has a free slot: the server runs up to MAX_CONCURRENT_TASKS at once (default 2), each computer you add has its own limit (1–32). A schedule skips or queues a firing while its last task still runs.

        Server or your computers

        Add a Mac in Settings → Computers (it runs fleetpocket-node with its own Docker and agent keys, and only connects out), then choose per project or per task where a task runs.

        Budgets and costs

        Every run's tokens and cost: reported by Claude, estimated for Codex and Cursor (≈$0.84). Totals by project, agent and day with a CSV export. Monthly caps per project, schedule or overall hold new runs at 100%, with a notice at 80% and a one-tap "Start anyway".

        CI auto-fix

        Off until you turn it on, globally or per project. Then a failing check on a Jungle Monkey PR starts a minimal fix from the trimmed CI logs, up to a few attempts and within a budget. A fix that touches CI config or tests waits for your OK.

        MCP through the gateway

        MCP servers (Linear, Sentry, a database, any npm/pip package or image) per project or globally. The agent reaches them through a gateway that holds the credentials, enforces tool lists and call limits, and logs every call.

        Schedules and loops

        Start a task on a timetable with its own budget, or let a task loop until your check (npm test, go test ./...) passes; the PR or local branch arrives once, at the end.

        Subagents

        Claude hands exploring and test runs to subagents on a cheap model, and a reviewer checks the diff. Their steps show as "via explorer", their cost counts toward the run, and the same sandbox policy applies. Claude only.

        § 03 · Observation

        See every agent work

        The canvas puts every running task on one board as a live node, whichever agent runs it: what it is doing now, its activity lanes, the files it read (R) and wrote (W), its steps, cost and time. Flip a node to a plain terminal for the same run as log lines. Answer a question, send a follow-up, or restore a checkpoint and let the agent continue from there.

        #42 · Retry webhook deliveriesClaude Code

        QUEUED

        waiting for a slot

        Touched

          STEP 0$0.000s

          § 04 · Security

          Sealed by design

          Whichever agent runs, it gets a folder, a prompt and its own provider key. No GitHub token, no Docker access, no host network. The GitHub and MCP credentials live outside its box. Tap a hop.

            System map: a sealed agent container holding only its own provider key; an egress proxy to allowed hosts; the MCP gateway to integrations; the publisher, which holds the GitHub token, to GitHub; the notifier to your phone. JUNGLE MONKEY SEALEDCONTAINER AGENTown key only WORKSPACEone repo EGRESS PROXYallowlist ✕ tracker.example.net ALLOWEDnpm · PyPI · API MCP GATEWAYcreds in memory TOOLSLinear · Sentry PUBLISHERholds GitHub key GITHUBpush · PR NOTIFIERTelegram bot YOUR PHONEanswer · merge

            § 05 · Deployment

            Install

            Three ways in. Start on your laptop in a minute; move to a server when you want it running while you sleep.

            npx jungle-monkey
            • Works with repos on your disk and gives you a branch back; connect GitHub to push and open PRs.
            • Needs Docker and an agent login (Claude, Codex or Cursor).
            • Stops when the computer sleeps.

            Local mode guide (LOCAL.md) →

            Compare
            QuestionOn your computerOn a serverSelf-hosted
            Where it runsyour laptop or desktop, in your Dockera small Linux VPSwherever you put it
            GitHub needed?no: local repos get a branch back; connect GitHub for PRsyes (your own GitHub App)yes (your own GitHub App)
            Phone accessnot yet: use a server installyes: PWA and Telegramyes: PWA and Telegram
            Runs 24/7no, stops when the computer sleepsyesyes
            Setup timeabout a minuteabout 10 minutesas long as you like

            § 06 · Field access

            Steer from anywhere

            One app on every screen. Phone and Telegram need a server install.

            Web

            The task list and the open task side by side: start, answer, follow up, merge.

            Desktop canvas

            Every running task as a live node on one board, grouped by the machine it runs on.

            Phone PWA

            Laid out for one hand. With the Ideas board turned on, drop in ideas (voice notes with a speech-to-text provider) and turn a card into a task.

            Telegram

            An optional bot and Mini App that ping you when an agent asks something.