v0.1.0

Published: 2026-07-20

Downloads for v0.1.0

filename	sha512 hash	size
mur-0.1.0-darwin-aarch64	`4e54cab636503a0b9443b2199c8d302753978bc5a36da429862de2af40413e7ae25bcdda899f5f4ca05f82140a42e61d56a9ec3066b52cc9653f9ab90f9fc6ad`	8.9MB
mur-0.1.0-darwin-x86_64	`7f2e198a21c59c497343623bca237a6721071c09321e31e9378b187036e7a1494fe7ebce2c59376010b135caaeb826864967ec95e36b911ee3b601df180b6ef8`	12MB
mur-0.1.0-linux-x86_64	`d777e23eff31faa639d2eafa5184344c9a5914bc97201335480dbe754b19fadd37cd00689f9c01bb48d0fd228fda56e16a43c74670cb2f9dca00f04a818464d3`	14MB

Changes since v0.0.0

Features

Package an agent into a single portable file, with everything it can access declared up front and nothing enabled by default.
Reproducible installs: capsules pin their versions and run the same way later.
Publish and install agents through a GitHub-backed registry, with latest / stable / edge release channels.
A public catalog of shareable artifacts with in-CLI search.
Run agents in a secure sandbox with per-agent network and file-access permissions.
Declarative lifecycle: run an agent once or keep it resident, with task scope and exit behavior set in the manifest.
Built-in agent mode: agents reason in a loop, run shell commands, and create their own tools mid-session.
Connect to major AI providers out of the box: Anthropic, OpenAI, and DeepSeek.
Give agents an identity and let them talk to each other (agent-to-agent messaging).
Long-running, multi-task agents with streaming output and live attach.
Pause an agent to ask a human for input, then resume (human-in-the-loop).
Coding tools for agents: a Git tool and a file editor (read, write, search, replace).
Skills: package reusable skills as artifacts, or wrap an existing SKILL.md into one.
Scaffold a new agent automatically, including a first-run provider setup wizard.
Automatic context compaction so long-running agents don't run out of room.
Persistent memory so agents remember prior work across tasks.
Smarter, cheaper context handling: bounded tool output, redundant-call detection, and stateful continuation between model turns.
Inspect exactly what an agent did through per-session traces and a trace CLI.
Monitor agents live with Grafana / OpenTelemetry dashboards.
Evaluate and score agent runs, with a dedicated eval CLI.

Other

Platform-wide security hardening: credential stripping, HTTPS-only pinning, install-time integrity checks, and signed session checkpoints.
Kernel-level sandbox enforcement, backed by the operating system.
Consistent, extensible lifecycle events for agents.
Shared vocabulary for agent state: sessions, tasks, turns, memory, and checkpoints.
Versioned internal interfaces with a documented policy for safe evolution.
Turn risky features on or off behind beta flags.