Harness 0.1
RU

Harness 0.1.0 · first public release

An agent on your server

A daemon with memory, tools and schedules. Thin clients, any model, your own files.

What is in 0.1.0 Capabilities
Harness chat: conversation, tool cards and the inspector

A screenshot of the real client. The data is a demo profile.

Version 0.1.0macOS · LinuxRust · SQLite · WebSocket

One daemon, many surfaces

The core does not know who connected. Logic lives in the daemon; clients only speak the protocol.

  1. 01

    Client

    The app, browser or CLI sends a message over WebSocket JSON-RPC.

  2. 02

    Context

    The daemon assembles prompt, memory, skills and the agent profile. Inside a session the context only grows.

  3. 03

    Policy

    Every tool passes allow, ask or deny. Anything unknown is denied.

  4. 04

    Turn

    The model answers, the daemon runs tool calls, counts budgets and can stop.

  5. 05

    Journal

    Every action is an append-only event. Search and projections are rebuilt from it.

Not a prompt in a chat, a working loop

The agent sees files, remembers decisions, reaches MCP servers and can work without you.

  • 01

    24/7 on your server

    systemd or launchd, auto-start, survives a closed laptop.

  • 02

    Agents and projects

    Own persona, tools, model and limits. A project is a folder and a default agent.

  • 03

    Memory and skills

    Markdown and YAML: diffable, portable, readable by humans.

  • 04

    Any model

    Anthropic, OpenAI, Ollama, local OpenAI-compatible servers. Fallback when a provider fails.

  • 05

    MCP and tools

    MCP client and server, install catalog, files and shell under policy.

  • 06

    Voice

    Whisper for input, say or espeak for speech. No cloud required.

  • 07

    Policies and approvals

    Deny-by-default, approvals in the chat feed, read-only token for observation.

  • 08

    Evals and tests

    YAML scenarios, redacted datasets, 146 tests and clippy in CI.

Different roles, one infrastructure

The reviewer gets no write tools. The scout can only read and search. Model, limits and working directory belong to the agent; memory and skills are isolated.

  • A profile is a YAML file, with JSON export and import
  • A single chat can override the model
  • Templates: code, review, research, ops
Agent profiles: Harness, Reviewer and Scout with models and tool sets

While you sleep

Cron jobs start a fresh session with the right agent and deliver the result to a file, stdout or a webhook. The schedule lives in config, not in a database.

  • Morning digest across projects
  • Nightly backup with an archive check
  • Delivery: file, stdout, webhook
Scheduled jobs: a morning digest at 8:00 and a nightly backup at 3:00

Two themes. One client.

Dark and light, switched without a reload. The same web client runs inside the desktop app and in a browser.

The same chat in the light theme

Security

Denied first, allowed by exception

Rules are computed in code, not in the prompt. The model cannot ask for more than it was granted.

  • deny: ~/.ssh, ~/.aws, sudo, rm -rf
  • ask: shell and unknown tools
  • allow: read, search, memory, skills
  • Admin and read-only tokens; secrets only in env and Keychain
  • local or docker sandbox without network
  • Audit and backups that contain no secrets
The Policy page: denials, confirmations and allowances

What works and what does not

This is a personal project, so the status has no asterisks.

Done

  • ✓ Release 0.1.0: core, memory, search, subagents, scheduler, MCP, voice, evals
  • ✓ macOS app, web client and CLI on one protocol
  • ✓ Server archives for Linux x86_64 and aarch64, run in containers

In progress

  • – Windows runtime builds but has not run on real hardware
  • – No notarization and no auto-update yet
  • – Code mode is limited to batch calls
What shipped in 0.1.0 →

It lives on your server

macOS app, Linux daemon, web client. One protocol, your files, your models.

harness.progdes.ru