Harness engineering,
from a window, not a dotfile
What an agent can do is decided by its harness: its persona, its tools and its memory. AgentsRoom turns them into three separate controls you can mix per agent and per task.
A toolkit bundles skills, MCP servers, saved connections and environment variables under one name. Attach it to an agent, a single ticket, a team step, a project or a client folder. Edit it once and every agent that uses it gets the change at its next launch, on Claude Code, Codex and every other agent CLI.
- scan-documents
- summarize
- notes (MCP)
- staging-db
- Claude Code
- Codex
- Antigravity
- Cursor
One research toolkit, attached to a ticket: the agent gets its skills, its MCP server and its connection at launch, and reads the project memory without writing to it.
Harness engineering is the work of shaping what surrounds a coding agent: which instructions it reads, which tools it can call, which knowledge it keeps. Most of it lives in files today. A CLAUDE.md here, an mcp.json there, a skills folder per CLI, environment variables in a shell profile, and none of it travels to the next agent or the next task without being copied by hand.
In AgentsRoom an agent is three things you set separately. The persona says who it is. Its toolkits say what it can use: skills, MCP servers, saved SSH, database and dev command connections, environment variables. The memory setting says whether it reads and writes the shared project memory. Change one without touching the other two.
A toolkit is attached by reference, never copied. A research toolkit made of a document scanning skill, a summarizing skill and a connection to your notes app can equip a reviewer agent, a single ticket and a whole client folder at once, and an improvement to it reaches all of them.
Three controls instead of one agent definition
Give a task capabilities without a persona, keep a persona without memory, or reuse the same tools under different roles.
Skills, MCP servers and connections in one toolkit
Pick skills from your library, saved SSH, database and dev command connections, paste the MCP servers JSON you already use and add environment variables. Credentials stay in the Secret Manager as {{secret:NAME}} references, resolved on your computer at launch.
Attached by reference
Agents, tickets and settings store the toolkit, not a copy of its content. Edit the toolkit once and every agent that uses it gets the new skills and servers at its next launch.
Every level you already use
Attach a toolkit to one agent, a saved agent, a single backlog ticket, a step of an agent team, a whole project, or a client folder whose sub-folders add their own toolkits.
Capabilities for one task
A ticket can bring its own toolkits without changing its agent. They add to the agent's own, or replace everything (agent, project and folder) when the ticket's skills mode is Replace.
Memory, separate from the persona
Set project memory to Off, Read only or Read and write per agent, and override it for one ticket. Read only keeps the knowledge base available without letting a review agent rewrite it.
Every agent CLI, desktop and phone
Skills reach every CLI as text and, on Claude Code and Antigravity, as native skills. MCP servers are written in each CLI's own format. Create and attach toolkits from the desktop, the mobile app, or let an agent do it with the AgentsRoom tools.
Build a toolkit in under a minute
No file to edit and nothing to restart by hand.
- 1
Open the library
Click the briefcase next to Agents in the title bar, or Manage next to any Toolkits field, then New toolkit.
- 2
Fill it
Name it, pick its skills and connections, paste its MCP servers and add its environment variables. Use {{secret:NAME}} for a token.
- 3
Attach it
In Edit agent > Advanced options > Toolkits, on a ticket, a team step, Project settings > Capabilities or Folder settings > Knowledge.
- 4
Choose the memory
Right below the toolkits, set Project memory to Off, Read only or Read and write, or leave it inherited.
- 5
Launch
The next launch of the agent expands the toolkit into its CLI: skills, servers, connections and variables, nothing written in your repository.
Why separate persona, tools and memory
Until now an agent definition carried everything at once. Reusing a set of skills and servers with a different role meant rebuilding it, and the only way to keep an agent away from the project memory was to strip its persona with the minimal system prompt.
Separated, the three controls answer real situations: a reviewer that reads the memory but never writes it, a one-off research ticket that needs a notes server its agent does not have, a client folder whose every agent must reach the same staging database, a team step that uses a browser toolkit only for its own turn.
It is also the entry point to harness engineering for people who do not want to maintain dotfiles: every choice is visible in the agent window, travels with your account to your other computers, and can be audited at a glance.
Good to know
- A change applies at the agent's next launch. A session that is already running keeps what it started with.
- A project skill or connection inside a toolkit only applies in its own project. Account skills apply everywhere.
- Toolkits belong to your account. In a team shared with colleagues, each person's launch only resolves their own toolkits.
- A team step has toolkits but no memory setting: use the step's Read only switch for a step that must not write anything.
- Up to 20 toolkits per agent, ticket, step, project or folder, and 200 per account.
- An agent running on a remote machine over SSH gets the toolkit's skills, but not its MCP servers.
Frequently asked questions
What is harness engineering?
It is the practice of shaping the environment of a coding agent rather than its prompt alone: the instructions it loads, the tools and MCP servers it can call, the knowledge it reads and writes. AgentsRoom exposes those pieces as three controls (persona, toolkits, memory) that you set per agent and per task.
Can I reuse the same skills and MCP servers on several agents?
Yes. Put them in a toolkit and attach it to each agent, or to the project or folder so every agent there gets it. Editing the toolkit updates all of them at their next launch.
Can a single ticket get an MCP server its agent does not have?
Yes. Attach a toolkit to the ticket. The run gets it on top of the agent's own toolkits, and the agent's saved configuration is not changed.
How do I stop an agent from writing to the project memory without losing its persona?
Set its Project memory to Read only, or Off to hide the memory completely. The persona and the tools stay as they are. You can also set it on one ticket only.
Does it work with Codex and the other agent CLIs, or only Claude Code?
Every agent CLI AgentsRoom launches. Skills arrive as text everywhere and as native skills on Claude Code and Antigravity, MCP servers are written in each CLI's own configuration format, and connections and variables follow the agent whatever its CLI.
Where are my tokens stored?
Never in the toolkit itself. A toolkit holds {{secret:NAME}} references to the Secret Manager, and the value is resolved on your computer when the agent starts.
Can my agents create toolkits for me?
Yes, when you ask them. The AgentsRoom tools include toolkits_list, toolkits_get, toolkits_save and toolkits_delete, and agents_save, backlog tickets, team steps and project or folder settings accept a list of toolkits. Agents are told never to create one to give themselves access.
You may also like
Multi-Agent Workflow • Handoff • Feedback Loop
AgentsRoom Agent Teams chains AI coding agents like a real engineering team. Visual workflow editor (n8n-style canvas) to wire Dev, QA, PM, Security and DevOps roles. Automatic handoff with feature summary, git diff, risks and test hints. Machine-verified quality gates (your test suite decides, not the agent), parallel QA and Security review branches, conditional edges, feedback loops, max-cycles guard. Choose per team how the roles run: one agent that morphs from role to role and keeps the whole context, or one live agent per role, messaging each other in real time. The multi-agent orchestrator your agents were missing, on Claude, Codex, GitHub Copilot CLI, Cursor and 13 other agent CLIs.
MCP servers
Declare an MCP server once and hand it to Claude Code, Codex and every agent CLI. Pick it from the official MCP registry, keep keys in the Secret Manager, and sign in to OAuth servers like Notion or Linear with one Connect button.
Project Memory
Give your AI coding agents a long-term memory. Decisions, architecture and pitfalls persist across sessions and are shared with everyone on the project. Free, unlimited.
Customize Agents
Pick from 14 built-in agent roles, add specialist agents from The Agency, or build your own custom AI coding agent with its own system prompt. Then share your agents with your team through the chat, and publish them to the community so everyone levels up together.
Secret Manager
Store API keys, tokens and passwords in your OS keychain, reference them as {ref} in dev commands and agent environments, and let AgentsRoom resolve them at launch. Agents see the names, never the values, and nothing is ever synced to a server.
Equip your agents once
Free on macOS, Windows and Linux, with the mobile app to manage toolkits from your phone.
Companion app: monitor your agents on the go
Bring your own: Claude, Codex, Antigravity CLI, or other AI provider.
Push bugs and requests straight to your public backlog.