Harness field

Same job. Different runtime. Different machine.

A harness is the loop around a model: entry, tools, memory, policy, persistence. This page teaches the catalog as structure. It does not rank these products. Public Index ranks stay on the labeled Terminal Core preview seed.

Browse the field Back to the index

Nine runtimes. Zero public scores.

Each row is identity plus a teaching structure. Status is the catalog ladder, not a rank. Jump to a breakdown, or keep scrolling.

External catalog only. Synthetic preview harnesses used by Playground are omitted. No Index ranks on this table.
Harness	Org	Status	Eval contract	Structure
Pi pi	Earendil Works	catalog-only	pi --print --mode json	Read
oh-my-pi oh-my-pi	can1357	catalog-only	unconfirmed	Read
Grok CLI grok-cli	superagent-ai	adapter-stub	grok --prompt "<task>" --directory <workdir> --format json	Read
Grok Build grok-build	xAI	catalog-only	grok --single --output-format streaming-json --always-approve	Read
Claude Code claude-code	Anthropic	catalog-only	claude -p --output-format json	Read
Codex CLI codex	OpenAI	catalog-only	codex exec --json	Read
OpenCode opencode	OpenCode	catalog-only	opencode run --format json	Read
Hermes Agent hermes	Nous Research	catalog-only	first-party headless argv unconfirmed	Read
Gemini CLI gemini-cli	Google	catalog-only	gemini -p --output-format json	Read

Earendil Works

A small kernel. Print, tools, sessions. Almost nothing else.

Pi. Pi is the upstream coding-agent kernel HI brands. Catalog it as Pi, not as HI, and not as the oh-my-pi fork.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-13

Eval contract

pi --print --mode json

Harbor already runs this with an explicit session directory. No public score until a pinned package reproduces twice.

toolsheadlesssessions

01
Entry
Interactive TUI or non-interactive print mode.
02
Loop
One agent turn. Tools fire until the print session stops.
03
Instructions
Layered AGENTS.md and CLAUDE.md from home, parents, and cwd.
04
Sessions
Transcripts under ~/.pi/agent/sessions/.
05
Policy
Project trust only. No built-in sandbox.

Teaching structure for Pi. Catalog notes, not a live product port and not an Index score.

Drawing structure

Do not confuse Pi with oh-my-pi or with HI. HI brands Pi and admits other CLIs on Herdr. This diagram is a teaching sketch from the catalog profile.

Open Ops row Upstream repo Docs

can1357

A Pi fork whose workers are copies of itself.

oh-my-pi. HI stole the TV-wall UI, not the director. Catalog omp so Index never treats vibe workers as foreign harnesses.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-13

Eval contract

unconfirmed

No Harbor installed-agent adapter. Do not assume omp --print matches pi --print.

toolssub agentssessions

01
Pi kernel
Same small loop family as upstream Pi, until proven otherwise.
02
Agent Hub
Director verbs: spawn, send, wait, kill, list.
03
Workers
Copies of oh-my-pi. Running, idle, parked, or aborted.
04
TV wall
Vibe-mode visual of those self copies.

Teaching structure for oh-my-pi. Catalog notes, not a live product port and not an Index score.

Drawing structure

Never score vibe_spawn of self as harness-versus-harness. HI harness mode maps those verbs onto real Herdr CLIs. That mapping is HI, not omp.

Open Ops row Upstream repo

superagent-ai

A community terminal agent with a documented headless contract.

Grok CLI. First external catalog example: tools, sub-agents, sandbox, MCP, skills, and grok --prompt --format json. Not affiliated with xAI.

Catalog status: adapter-stub
Adapter: stub · 0.1.0-stub
Reviewed: 2026-08-12

Eval contract

grok --prompt "<task>" --directory <workdir> --format json

Adapter is a stub. Pin grok-dev before any scored-preview climb. Interactive TUI is out of scope for Index runs.

toolssub agentssandboxheadlessbrowsercomputer usemcpskillssessionsremote controlverifyhooks

01
Entry
TUI, headless --prompt, Telegram bridge, or daemon.
02
Agent core
streamText tool loop, compaction, delegations.
03
Tool plane
bash, files, grep, search, media, computer-use, LSP.
04
Extensions
Hooks, MCP, skills, AGENTS.md, custom sub-agents.
05
Persistence
SQLite transcripts, usage, workspace trust.

Teaching structure for Grok CLI. Catalog notes, not a live product port and not an Index score.

Drawing structure

Community-built. Not affiliated with, endorsed by, or sponsored by xAI. Grok is a trademark of xAI Corp. Do not merge this row with official Grok Build.

Open Ops row Upstream repo

xAI

Official xAI Grok Build. A different product from community grok-cli.

Grok Build. Two CLIs share a name. This row is the official binary. hi kind grok currently documents the community CLI.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-13

Eval contract

grok --single --output-format streaming-json --always-approve

Harbor pins web search off because it bypasses environment isolation. Do not overload the grok kind.

toolsheadlesssessions

01
Headless
--single / -p with an explicit output format.
02
Approvals
--always-approve for unattended runs.
03
Sessions
Headless history under ~/.grok/sessions.
04
Isolation note
Web search is a server-side tool. Harbor turns it off.

Teaching structure for Grok Build. Catalog notes, not a live product port and not an Index score.

Drawing structure

This is the xAI product. Community grok-cli is a separate MIT CLI. An .env-upload incident on the community tool is not a score on this row.

Open Ops row Upstream repo Docs

Anthropic

The natural peer terminal agent. Cataloged without a frozen contract.

Claude Code. Daily-driver peer to Codex. Include it so the field is honest. Do not invent a score.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-12

Eval contract

claude -p --output-format json

Without --bare, print mode loads hooks, skills, plugins, MCP, and CLAUDE.md. --bare is recommended for CI. Index has not frozen either path.

toolssub agentsheadlessmcpskillssessionshooks

01
Print mode
claude -p / --print with text, json, or stream-json.
02
Auto context
Hooks, skills, plugins, MCP, memory, CLAUDE.md unless --bare.
03
Tools
bash, read, edit, glob, grep, plus product sub-agents.
04
Sessions
Interactive and resume workflows. Eval mapping still TBD.

Teaching structure for Claude Code. Catalog notes, not a live product port and not an Index score.

Drawing structure

Proprietary. A widely used -p flag is not an Index freeze. Evals that omit --bare will not be comparable to the future print-mode default.

Open Ops row Upstream repo Docs

OpenAI

Official non-interactive path is codex exec, not a guessed argv.

Codex CLI. Practitioner 1–2 with Claude Code. Stars underweight it. Catalog identity now; freeze later.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-13

Eval contract

codex exec --json

Default sandbox is read-only. Automation should pick workspace-write or danger-full-access. Harbor’s bypass flags are Harbor’s.

toolssandboxheadlesssessionsskills

01
codex exec
Progress on stderr. Final message on stdout.
02
JSONL
thread.started, turn.started, item.*, turn.completed.
03
Sandbox
Read-only by default. Explicit write modes for automation.
04
Resume
codex exec resume --last for multi-turn suites.

Teaching structure for Codex CLI. Catalog notes, not a live product port and not an Index score.

Drawing structure

Do not treat Harbor’s --dangerously-bypass-approvals-and-sandbox as the Index freeze. First-party user-count claims are not Index inputs.

Open Ops row Upstream repo Docs

OpenCode

Largest MIT coding CLI in the census, distant third on X.

OpenCode. Big OSS with a documented run path. Catalog the gap between stars and daily-driver rank without scoring it.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-13

Eval contract

opencode run --format json

Official docs list --auto for permission auto-approve. Harbor’s --dangerously-skip-permissions was not on the fetched CLI page.

toolsheadlesssessions

01
opencode run
Scriptable path. --format default or json.
02
Serve / attach
TUI is a client of opencode serve.
03
JSON events
text, reasoning, tool_use, step_start, step_finish, error.
04
Permissions
Prefer documented --auto until a suite needs more.

Teaching structure for OpenCode. Catalog notes, not a live product port and not an Index score.

Drawing structure

Star count is not a quality or Index rank. Reconstructing a missing user-turn event, if needed, is an adapter note, not a silent rewrite of the product.

Open Ops row Upstream repo Docs

Nous Research

A large public agent with a Harbor adapter and no Index freeze.

Hermes Agent. Identity only. Star quality is unaudited. Do not invent a first-party headless flag.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-13

Eval contract

first-party headless argv unconfirmed

README documents an interactive CLI. hermes -z is a stricter one-shot. Harbor does not use -z. Confirm live --help before leaving catalog-only.

toolssessions

01
hermes CLI
Interactive entry is documented. Eval entry is not frozen.
02
One-shot
hermes -z: final reply only on stdout.
03
Harbor adapter
Exists. Not an Index freeze.
04
Sessions
Trajectory may depend on session export.

Teaching structure for Hermes Agent. Catalog notes, not a live product port and not an Index score.

Drawing structure

Do not publish an Index score from Harbor’s adapter alone. Star count, forks, and issue totals are not quality.

Open Ops row Upstream repo

Google

Eval-able, Apache-2.0, and not an hi kind yet.

Gemini CLI. Meets the install-plus-headless bar. Station admit will refuse it until a kind exists.

Catalog status: catalog-only
Adapter: planned · 0.0.0-planned
Reviewed: 2026-08-13

Eval contract

gemini -p --output-format json

Harbor has an adapter. HI cannot admit this kind until a kind is added.

toolsheadless

01
gemini -p
Non-interactive prompt entry.
02
JSON output
--output-format json for scripted runs.
03
hi kind
None. Identity-only until the kind registry grows.

Teaching structure for Gemini CLI. Catalog notes, not a live product port and not an Index score.

Drawing structure

Large OSS CLI. Still no public score without a frozen suite, and Station cannot admit it until someone adds a kind.

Open Ops row Upstream repo