exPardus

Open source

Built in the open.

Tooling we built for ourselves and published. All three flagship projects run on our own repositories; the numbers below were measured against the code, not estimated.

github.com/exPardusFive public repositoriesMeasured 10 Sep 2026

Our engineering runs on Claude Code. When you run several coding agents at once across several projects, three things go wrong: you lose track of them, each one gets stuck in its own dead ends, and sometimes you want a second model's hands on a task. claude-fleet is our answer to the first, cc-oracle to the second, multi-codex to the third. All three are plain, dependency-free tools that stay out of the way.

01

claude-fleet

Run a whole team of Claude Code sessions from one seat.MIT · v0.2.0

claude-fleet is an orchestration layer on top of Claude Code's own background agents. One manager session spawns workers across every project on your machine, steers them mid-turn without attaching a terminal, and hands the fleet off cleanly when it restarts. Workers are durable Claude Code sessions on disk: they survive crashes, reboots and the manager's own death.

The whole thing is a single-file, stdlib-only Python CLI plus a few hooks. State is plain files; every surface — statusline, slash commands, the manager session — is a disposable view over the same files. Nothing is pushed into sessions that did not ask for it.

What it does06

  • Mid-turn steering

    Drop a message into a running worker's mailbox; it is injected at the next tool boundary. No attach required.

  • Token ceilings, enforced fleet-side

    A worker that would blow its cap refuses to continue — checked before every turn, not after the bill.

  • Respawn with journal continuity

    Give a worker a fresh context while carrying its journal and mailbox forward. The context-reset lever for long campaigns.

  • Park on usage limits, resume later

    A worker that hits a plan limit parks itself with a reset horizon instead of dying silently, and is relaunched when the window passes.

  • Durable manager identity

    Boot-claim, heartbeat and hand-off protocol so exactly one manager owns the fleet across restarts — and can pass the baton.

  • fleet doctor

    28 health checks in one command: registry, hook wiring, stale sessions, orphaned mailboxes, version pins. Report-only by default.

fleet is built and verified on Windows and Linux; on Windows the shipped shim needs Python 3.13. macOS shares the POSIX backend but has not been receipted yet, and the repo's launch-readiness document lists everything that still blocks a first-time user.

02

cc-oracle

A Claude Code plugin that stops your session from flailing.MIT · v0.3.0

Every agentic coding session hits a wall sometimes. Left alone, a model keeps trying, and each attempt adds noise to its own context — the tenth attempt is being made by a model reasoning over nine failures. cc-oracle changes the default: when a session notices it is unsure, it sends a structured brief to a fresh-context subagent on the best available model, restricted to read-only tools, and implements the diagnosis and plan it gets back.

This makes a cheap day-to-day model plus an on-demand expensive consultant a workable setup. It ships as precompiled binaries for Windows, macOS and Linux, needs no Python or Node, and fails open — a hook must never wedge a session.

What it does04

  • Doctrine

    A standing instruction injected at session start: uncertainty is a signal, not a failure — consult the oracle before you pollute your context.

  • A read-only oracle

    A subagent with Read, Grep and Glob only. It cannot change anything, so a consult is always safe, and it answers in a fixed shape: diagnosis, plan, pitfalls.

  • A safety net for forgetting

    If the model says "I'm not sure" and did not consult, the turn is blocked once with a nudge quoting its own words. Conservative, and fails open.

  • A trigger for silent stalling

    Three consecutive failing tool calls earn the same nudge. The threshold was set by measuring 1,741 real turns.

What the benchmark shows and does not: across 104 runs on three models, the plugin cost nothing when nothing stalled and its Stop hook produced no false positive in 398 turns. It does not measure rescue — no run got stuck, so no consult happened. That claim rests on real work, and the repo says so.

03

multi-codex

One shell. Independent workers.v0.1.0

multi-codex gives a coordinating Codex or Claude Code session independent Codex workers through the shell. Hand off a focused task, keep working, collect the answer later. Each worker is its own codex exec process outside the coordinator's subagent pool: it starts fresh, writes its answer to a file and exits.

It is deliberately small enough to read in one sitting: 239 lines of Bash, five commands, and a .mcx/ folder of plain files for prompts, state, logs and answers. Workers may use native Codex subagents but cannot launch more workers, so a fleet cannot spawn recursively by accident.

What it does05

  • Fire and forget

    spawn returns a worker id immediately; the worker survives the shell that launched it closing.

  • Explicit context

    You pass the task and everything the worker needs. No parent transcript is copied, so nothing leaks by accident.

  • Steer mid-course

    steer interrupts the run and resumes the same saved conversation with a new instruction — even after the worker has finished.

  • Cost-conscious defaults

    Workers run on a small model at medium reasoning unless you say otherwise; model and effort can be set per worker.

  • Inspectable state

    Prompts, status, logs and answers are files in .mcx/. No scheduler, no service, no database.

First released on 10 September 2026 and verified against Codex CLI 0.153.4 and Claude Code 2.1.267; your account's usage and rate limits still apply, and a steer is a resume of the saved conversation, not live message injection.

04

Also public

Smaller things we released along the way
Why we publish
We open-source the tools that made us faster because tools like these get better with more hands on them — and because a company that says it builds well should be able to show its work.

Use it, break it, tell us

Issues and pull requests are welcome on GitHub. Questions about how we use these tools in our own work? Write to us.

github.com/exPardus (opens in a new tab)Write to us