agent-rosetta

Agent Harness
Layers

Every agent harness implements the same five extension layers — only the names differ, and occasionally conflict. Instructions persist, skills teach, MCP connects, bundles ship, subagents delegate. Tap a layer heading for detail.

Where to start

Start from what you're trying to change, not from the feature name.

"Follow these rules in this project, every time"
Instructions
CLAUDE.md AGENTS.md SOUL.md .goosehints
"Do this kind of task my way, when it comes up"
Skill
SKILL.md folder
"Read/act on my email, files, tickets, database…"
MCP server
Connector Extension app
"Give my whole team this setup in one install"
Plugin / Extension / Package
bundle of all the above
"Split this into parallel or isolated work"
Subagents
delegated agent instances
Show
← scroll → on mobile · tap a column header for the full explanation · use the picker to focus one harness

The five layers, explained

Same architecture everywhere, and the order is the stack: rules → know-how → access → packaging → parallelism. Each layer only earns its place when the one before it isn't enough.

Terminology traps

  • "Plugin" has several meanings across these tools. Claude, Codex, Goose, Grok Build, Cursor, and Copilot use it for installable bundles. Hermes and OpenCode use it for executable code extensions. OpenClaw supports both native code plugins and imported content bundles, while Osaurus imports Claude plugins into its own feature set. Check the manifest and payload, not just the name.
  • "Extension" flips meaning too. In Goose, an extension is an MCP server (tool access). In Gemini CLI, an extension is a full agent bundle. In Zed it is an editor bundle that can carry an MCP server but has no documented skill or instruction payload. Do not confuse current Copilot plugins with the retired Copilot Extensions GitHub Apps.
  • "Skills" share a core format. Every harness here uses the same Agent Skills / SKILL.md core format. The wrinkles are discovery paths, precedence, trust, and distribution: several tools read Claude-compatible and vendor-neutral .agents/skills directories alongside their own.
  • "Connector" is mostly Claude-speak for a packaged remote MCP server — though the word shows up generically elsewhere to mean any MCP integration.
  • "Bot" is not a subagent. A bot is a named teammate that stays. Hermes Bot Mode is a roster over profiles; Osaurus Agents are saved configs; Grok Bot is a separate always-on product from Grok Build. This column is a throwaway worker: spawn, return a summary, done.

The layer underneath: MCP

The Model Context Protocol (open standard, released by Anthropic in Nov 2024 and now governed by the Linux Foundation's Agentic AI Foundation) is the shared plumbing for the "tool access" row — a JSON-RPC protocol where a server publishes tools any compatible agent can discover and call. This is why the ecosystem interoperates: a Goose extension, a Claude connector, a Hermes MCP entry, and an MCP server inside a Codex plugin can be the exact same server. Osaurus demonstrates the two-way nature of the standard: it consumes remote MCP servers and exposes its own local tools as an MCP server that other harnesses can connect to — so your Osaurus toolbox can serve Claude Desktop. Skills (SKILL.md / agentskills.io) and instruction files (AGENTS.md) are similarly portable open standards, and since October 2026 all three sit under that same vendor-neutral foundation, alongside goose. The names are vendor branding; the files and protocols underneath increasingly are not.

Portability

Three things travel between harnesses as open standards — AGENTS.md instructions, SKILL.md skills, and MCP servers, all now governed by the vendor-neutral Agentic AI Foundation — and several tools read each other's directories outright, so one folder can serve many. Hooks travel by convention instead: there is no hooks standard, but Copilot, Cursor and Grok load Claude Code's hook files, though a script may need small changes to its input handling. Everything else is stranded. Tap any row for the exact paths.

← scroll → on mobile · tap a row for exact paths and what doesn't port

paths are what each project's current docs specify — verify before pointing one folder at several tools

Permissions
& Autonomy

Every harness names this differently — approval modes, allowlists, sandboxes, trust levels — so this table reduces it to two things: whether it edits your files and whether it runs your shell before checking with you. Then what contains it, and where you change it. Tap any row for the default in full, approval granularity, and unattended behaviour.

← scroll → on mobile · tap a row for the full picture and provider docs

defaults affect what can run without approval and change between releases — confirm against your installed version, not this table

ADEs
Agentic Development Environments

A harness runs one agent. An ADE runs several side by side, usually from different vendors, each in its own git worktree, and handles the work around them: watching, reviewing the diff, opening the PR. They differ most in how they drive each agent and how much of the review they handle. Most reuse the sign-ins your CLIs already have, and few sandbox anything on your machine. Tap a row for remote access, license, maturity and extras.

I'm on
← scroll → on mobile · tap a row for remote access, license, maturity, extras and sources

Provider
Compatibility Matrix

This table shows which tools will take a ChatGPT, Claude, Gemini, Copilot, Grok or OpenRouter account, which want an API key instead, and which won't connect at all. Turn off the credentials you don't have and the rows narrow to what you can use. When a tool reaches a provider through another agent, the cell names that agent. Tap a row for details and the vendor's docs.

I have
Show
← scroll → on mobile · tap a row for connection details · toggle credentials to filter columns

provider support changes frequently — this is a point-in-time snapshot; consult each tool's linked documentation before relying on it

Infrastructure
Your Agent Can Drive

The hosts and databases behind a vibe-coded project increasingly ship their own MCP servers, which means your agent can reach production directly — every one ships an official MCP server, and all but Render and the three search providers expose full control through it. Start by picking what you're building — a web app, a web app with a database — and the tables narrow to the platforms that cover it. Tap any row for detail.

I need
Show
← scroll → on mobile · a dash means the platform has no first-party answer — you'd bring your own

a tag under a cell is the catch: marketplace and partner mean someone else runs it, self-run means you operate the container, beta means not yet for production

…how much of it is free

Same rows, read as a starting cost. Everything above is available to a signed-up account with no payment method on file, except where this table says otherwise.

← scroll → on mobile · allowances are per month unless the cell says otherwise

a free tier is not the same shape everywhere: some are quotas that fail closed, some are credit that runs down, and one is a trial with a clock on it — the third column is where each one ends

tap any row above for its MCP server, riskiest operations, CLI and the rest of the scoping detail — reach is full control nearly everywhere, so check read-only modes, token scope and approval settings

What's new