AI coding tools compared for teams: harness, IDE, orchestrator
The short answer
If you want a local harness that splits, tests and merges work, pick CodeAF. If you want a small MIT base to extend yourself, pick Pi 1.0. If you want agents callable from CI without hosting anything, pick epho. If you run several CLI agents at once, pick Orca.
Drafted with AI from each tool’s own pages. Reviewed and approved by Pego.
Pego’s read
Drafted with AI from Pego’s own notes and the pages it covers. Reviewed and approved by Pego.
Skip to the evidenceThe CodeAF README opens with a number: “#1 on DeepSWE of ten coding harnesses on the same model, ahead of Claude Code, Codex, OpenCode, Kilo and DeepSeek’s own harness.” That is the sentence a comparison page trains you to read as an answer. The rest of this collection makes it useless as one. Herdr keeps terminal agents alive after you close the window. tgrep is a search index that GitHub Copilot CLI calls. epho is an HTTP endpoint that boots an agent in a cloud sandbox. None of them loses to CodeAF on DeepSWE, because none of them is in that race. So I would not pick a winner here. There is no best tool. There is the best tool for your need, and naming the need comes first.
Read the twelve items by the shape of the work and they split cleanly. CodeAF and Pi are harnesses you run on your own machine. tgrep is a layer underneath an agent, already integrated into GitHub Copilot CLI. epho is the opposite of local: you POST a prompt and the box is built for you. Kiro Crew is an orchestrator, with a crew of agents, approval gates and a dashboard. Herdr is a native window around terminals you already run, and Orca is a desk where several CLI agents work in parallel. Comparing these on one axis gets you nowhere, and the pages that sell one number are the ones that tell you least.
Operational detail is what earns trust
Two pages here earn their place by talking about the boring part. tgrep reports peak memory with private bytes kept apart from mapped pages, and says that indexing a 2 GiB file held 77.8 MiB private. That is the kind of detail I trust, because it costs something to publish and it is the number that decides whether the thing runs on a laptop. Pi 1.0 lists deferred tool loading and cache warming for Anthropic models among its release features. Those are the boring operational details that matter in production, and they are the kind of thing you only learn after running agents at scale.
The other end of the collection is a set of pages that make a claim and stop. Bullet says it is 30 to 60% faster than Claude Code and Codex on the same tasks, with no method and no task set behind the number. Kiro Crew does not name the agents it drives and does not state its licence, though the page says it grew out of an internal Amazon tool used by more than 39,000 engineers, which says the approval model has been worn in. epho says nothing about pricing, limits or self hosting, and its own API sends repos and a provider key on every request. That is not a reason to dismiss them. It is a reason to read a speed percentage as a claim and not a result.
The claim I would not take on faith
CodeAF is the one I would look at hardest. Building a subharness for reviewing pull requests as a specialist is a smart move, because review has a clear pass or fail and a harness can be judged on it. But the #1 on DeepSWE claim lives in CodeAF’s own README, and I would trust it only after running it myself. The page is honest about the rest: Apache 2.0, one small Go binary, and an early preview that is young and moving fast, with rough edges. That last line is worth more than the ranking.
Herdr is honest in the same way, and it costs the page its headline. Quitting or updating the app changes nothing, because the agents run in herdr-core, a small server on your machine. Then the page says that if the machine reboots or herdr-core stops, the programs stop too. Closing the window is not the hard part. The durable agent story is still a laptop that stays awake.
The desk where several agents run
Orca is the one that makes sense to me first. Running Claude Code, Codex and OpenCode side by side, each in an isolated worktree, solves the collision problem where it actually happens, which is the checkout and not the prompt. It is free, open source and runs on macOS, Windows and Linux. The rest is a very good developer experience. What the page does not mention is approval gates, which Kiro Crew puts in the loop, and unattended runs. So Orca fits a person at a desk watching several branches, not a crew working while nobody is there.
Put together, the collection argues against its own format. The tools that survive a second read are the ones that describe their own failure: tgrep’s memory numbers, Herdr’s reboot, CodeAF’s early preview, Orca’s worktrees with nothing said about approvals. The comparison worth making is not between products. It is between the shapes of work, and a team that cannot name its shape will pick whatever carries the best number on the page.
Pick it if…
- CodeAF: Open-Source Software factory for Open Modelsyou hand many small tasks across projects to open models and want the checking and merging handled
- Pi 1.0 | Earendilyou want a small MIT-licensed harness to extend yourself, not a fixed workflow
- Trigram-indexed grepyour agent loops stall on grepping large repos and you want a local index under them
- epho — agents as an apiyou want agents callable from a script or CI without running any infrastructure
- Bulletyou want to pressure-test the speed claim against your own harness before switching
- Kiro Crewyou want a crew of coding agents running unattended with approval gates and a dashboard
- Port22you want to check on and drive coding agents from a phone while away from the keyboard
- Aurayou want coding agents driven from an IDE with git and intent tracking in the loop
- v0 Design Systems 2.0you want design system work on a shared canvas rather than in a terminal harness
- Herdr: made for your agents, remade by themyou want a GUI around terminal agents you already run, not another harness
- Orca - Agent Development Environment (ADE)you run several CLI agents at once and want each in its own worktree
- bbyou care more about an editor and agent sharing one codebase than about a feature list
How they differ
The table scrolls sideways on a narrow screen.
| Tool | Open source | Where it runs | Primary interface | Runs other agents | Unattended runs |
|---|---|---|---|---|---|
| CodeAF: Open-Source Software factory for Open Modelsgithub.com | Yes, Apache 2.0 | Local machine, one small Go binary | One window, projects opened as tabs | Its own subharnesses, other agents not stated | Yes, merges what passes and asks only when it must |
| Pi 1.0 | Earendilearendil.com | Yes, MIT | Installed on the local machine by script or npm | Terminal UI, full-screen by default | MCP support, other coding agents not stated | Long-running tasks only through the separate Pi Durable package |
| Trigram-indexed grepgithub.com | Yes, MIT | Local machine, client and server | Command line, plus MCP search tools for Codex or pi | Not stated | Not stated |
| epho — agents as an apiepho.io | Not stated | Cloud sandboxes run by epho, not the local machine | One HTTP endpoint, JSON in, SSE stream back | Yes, Claude Code, Codex or OpenCode, picked by a field | Yes, async endpoint plus webhook, events replayed on reconnect |
| Bulletcodewithbullet.com | Not stated | Not stated | Not stated | Not stated | Not stated |
| Kiro Crewkiro.dev | Yes, open-sourced by AWS, licence not stated | Not stated | Live dashboard for watching the agents work | Yes, a crew of coding agents, which ones not stated | Yes, unattended, with approval gates in the loop |
| Port22tryport22.com | Not stated | Not stated | Phone | Yes, Claude Code, Codex and other coding agents | Not stated |
| Auraauravcs.com | Yes | Not stated | IDE | AI coding agents, which ones not stated | Not stated |
| v0 Design Systems 2.0v0.dev | Not stated | Not stated | One collaborative canvas | Not stated | Not stated |
| Herdr: made for your agents, remade by themherdr.dev | Yes, Apache 2.0 | Local machine, plus other machines over SSH | Native app with terminal panes | Yes, anything that runs in a terminal | Yes, agents keep running after the app closes |
| Orca - Agent Development Environment (ADE)www.onorca.dev | Yes, free and open source | Desktop app on macOS, Windows and Linux | Desktop app with terminals, a built-in file editor and git tracking | Yes, Claude Code, Codex, Gemini, Cursor CLI and other CLI agents | Not stated |
| bbgetbb.app | Not stated | Not stated | IDE | Not stated | Not stated |
Where they sit
Where it runs: Local machine → Cloud service
Interface: Text only → Full GUI
The verdicts
CodeAF: Open-Source Software factory for Open Models
A local coding harness that splits work into tasks, checks them and merges what passes, from one window.
Built for open models on the local machine, and the useful part is the middle: it splits work, runs the parts that cannot collide, tests and merges what passes. That suits a team handing over many small tasks across projects. The catch is age: an early preview with rough edges, and the #1 benchmark claims come from its own README.
Pi 1.0 | Earendil
A minimal, extensible terminal agent harness, with a separate experimental package for long-running agentic apps.
Pi keeps its discipline: a small terminal harness, MIT licensed, one curl to install, and features added only once they prove themselves. That suits a team wanting a base to extend rather than a fixed workflow. The catch: the long-running story sits in Pi Durable, a separate experimental package, and nothing here claims task splitting or merging. CodeAF ships that middle part. Pi leaves it to the team.
Trigram-indexed grep
A local index and search server that agents query instead of scanning the filesystem on every grep.
Where CodeAF and Pi are harnesses, this is the search layer an agent calls: tgrep sits inside GitHub Copilot CLI. The reporting is careful about memory, private bytes kept apart from mapped pages. That suits a team whose agent loops stall on grepping large repos. The catch is that it is a dependency, not a workspace.
epho — agents as an api
A cloud API that boots a sandbox with Claude Code, Codex or OpenCode and streams the run back.
The shape is the point: one endpoint, the harness swapped by a field, the provider key passed through so the provider bill stays with the caller. That suits a team wiring agents into CI, cron or a script without hosting anything. The catch is the hosted middle layer: repos and key travel on every request, and the page says nothing about pricing, limits or self-hosting. CodeAF and Pi stay on the local machine.
Bullet
A coding agent whose only stated pitch is running 30 to 60% faster than Claude Code and Codex.
The pitch is one number: 30 to 60% faster than Claude Code and Codex on the same tasks. That is a claim on the page, not a result anyone can check, and a speed leaderboard is only as good as its tasks. A team already swapping harnesses per request, as epho allows, needs the method first. The catch: licence, interface and where it runs are not stated.
Kiro Crew
An AWS open-source orchestrator that runs a crew of coding agents unattended, with approval gates and a live dashboard.
The shape is an orchestrator rather than a harness: a crew of agents, approval gates, persistent memory and a dashboard to watch them. The 39,000 engineer figure is internal Amazon use, which says the approval model has been worn in. The catch is what the page leaves out: which agents it drives, the licence, and where it runs. epho at least names its harnesses.
Port22
Drives Claude Code, Codex and other coding agents from a phone, for when a build breaks away from a keyboard.
The one clear fact is the phone: coding agents reachable from a handset when a build breaks and nobody is at a keyboard. That covers a gap the harnesses here leave open, since they assume a desk. The catch is the page. No licence, no price, no word on where the run happens. epho at least names its transport and its key handling.
Aura
An open-source IDE for driving AI coding agents, with git and intent tracking inside the loop.
The pitch is integration: git and intent tracking sit inside the agent loop instead of being bolted on. That suits a team that wants agents driven from an IDE rather than a terminal harness like Pi. The catch is the page. It is a single line, with no licence, no named agents and no word on where the run happens.
v0 Design Systems 2.0
An AI assistant that builds whole design systems and full-stack apps on one shared canvas.
The move from single screens to a whole design system on one shared canvas is the claim, and it puts the work where a designer and a developer can both see it. That suits a team producing many screens that must look the same. The catch is the page: no licence, no word on where the run happens, no price. Aura at least states it is open source.
Herdr: made for your agents, remade by them
A native desktop app for terminal coding agents, with a local server that keeps them running when the window closes.
Herdr is a native window around terminal agents, with a local server as the useful part: close the window and the panes stay alive. It suits a team that wants a GUI around agents it already runs, where Pi stays a terminal harness. The catch is durability: if herdr-core stops or the machine reboots, the programs stop, so the story still depends on a laptop that stays awake.
Orca - Agent Development Environment (ADE)
An agent development environment: many CLI agents in parallel, each in its own isolated worktree.
Orca's bet is the worktree: each CLI agent gets its own isolated copy of the repo, so several run at once without stepping on each other. Free, open source, and it stays on the desk. That suits a team already running Claude Code and Codex side by side. The catch: nothing on approval gates, which Kiro Crew puts in the loop, and nothing on unattended runs.
bb
An IDE pitched as proof that the editor and the agent building it can be one evolving codebase.
The claim is unusual: the editor and the agent working on it are the same codebase, so the tool is also its own demo. That is a statement about how software gets built, not a feature list. A team choosing a workspace gets nothing to compare on licence, where it runs, which agents it drives, or whether it runs unattended. Aura at least states it is open source.
Still open
- Bullet’s 30 to 60% speed claim has no stated method or benchmark.
- Kiro Crew does not state its licence, which agents it drives, or where it runs.
- epho states no pricing, limits or self-hosting option.
Sources: github.com ×2, auravcs.com, codewithbullet.com, earendil.com, epho.io, getbb.app, herdr.dev, kiro.dev, tryport22.com, v0.dev, www.onorca.dev.