v0.9.0 · MIT · Bun + Node + TypeScript

One harness. Every coding agent.
Portable discipline.

Ultimate Harness plans, launches, observes, verifies, and promotes agentic software-development work across Hermes, Codex, Hermes Proxy, and Oh-My-Pi — without losing your specs, sandbox boundaries, or audit trail.

Read the docs →Try the TUIGitHub
~/agent-work · zsh
# install
$ bun add -g @agenticengineeringagency/ultimate-harness

# initialize a project
$ uh init

# add adapters (hermes, codex, hermes-proxy, oh-my-pi)
$ uh adapter add codex

# propose, run, verify
$ uh propose my-mission --workflow research-docs --goal "Refactor X"
$ uh mission run my-mission --runtime codex
$ uh verify my-mission

# or just open Mission Control
$ uh tui

Adapters

hermesactive

Direct Hermes server runtime, pinned to ≥ 0.14.0.

codexactive

OpenAI Codex CLI transport with sandbox routing parity.

hermes-proxyactive

ToS-positioned subscription gateway (nous / openai / anthropic providers).

oh-my-piexperimental

Anthropic via Oh-My-Pi for ANTHROPIC_API_KEY routing.

Pipeline

Every mission moves through the same seven phases. Each phase is inspectable, replayable, and gated.

  1. 1
    research

    Inputs → narrative → adopt / reject / defer.

  2. 2
    spec

    Mission packet drafted; acceptance criteria locked.

  3. 3
    plan

    Workflow profile + runtime config + dispatch.

  4. 4
    sandbox

    git-worktree spun up; mission bound.

  5. 5
    run

    Adapter executes; events.ndjson tails live.

  6. 6
    verify

    Checks gate promotion; review notes attached.

  7. 7
    promote

    Sandbox → canonical; audit trail sealed.

What's in the box

Mission packets

Portable, versioned work requests with workflow profile and runtime config. One spec, every runtime.

Read more →

SDD + TDD + cross-runtime QA

Three composable discipline layers: structured acceptance_criteria, test-first diff gate, and `uh mission run-all` fan-out comparison.

Read more →

Sandbox isolation

git-worktree sandboxes by default; mission-bound, dirty-aware discard, ready for AgentFS swap-in.

Read more →

Live Mission Control TUI

Three-pane dashboard, mission drilldown with Code/Diff viewers, R-runs with live events.ndjson tail.

Read more →

Runtime adapter contract

Every adapter implements the same shape: dryRun, run, events emission, runtime-result schema.

Read more →

Verify and promote

Sandbox work becomes canonical work only after verification gates pass. Audit trail is built in.

Read more →

Audit-by-default

Append-only events.ndjson per mission; runtime-session.yaml summary; diff.patch captured per run.

Read more →

Built for agent-driven engineering.

A project should be able to use multiple coding-agent runtimes without losing its specifications, skills, workflow state, audit trail, sandbox boundaries, or human approval checkpoints. UH gives you the spec format, the dispatch table, the sandbox backend, the verification gate, and the live UI — all in one binary.

Product briefArchitecture