AgentsRoom MCP :
an IDE your agents can drive
AgentsRoom MCP exposes the AgentsRoom IDE to your AI coding agents through the Model Context Protocol. Four MCP servers (Backlog, Terminal Commands, Prompt Library, Browser) give Claude, Codex, GitHub Copilot CLI, Cursor and 13 other agent CLIs the ability to read, write and act on the workspace itself.
Most actions can stay 100% manual : you click, you drag, you type. The same actions can also be triggered by an agent through AgentsRoom MCP, end to end. AgentsRoom MCP is the unique edge that pushes AgentsRoom past a regular IDE and turns it into an agent-controllable workspace.
AgentsRoom MCP : an AI agent uses the Terminal Commands MCP server to start the dev server, the new tab shows up live in the Terminal Commands modal.
AgentsRoom MCP is a set of MCP servers that expose the AgentsRoom IDE to the agents running inside it. Until now, an AI coding agent could read your code, run shell commands and edit files. With AgentsRoom MCP, the same agent can also pilot the IDE itself : create backlog tickets, mark them done, save a dev command and start it, save a prompt to the library, open a browser tab and click around. The IDE becomes part of the agent's tool surface.
Concretely, AgentsRoom MCP ships four MCP servers today. Backlog MCP gives the agent CRUD over your project backlog and even lets it answer public backlog tickets directly to your users. Terminal Commands MCP lets the agent list, save and run dev commands : the agent can start the site dev server, the API, the worker, and any new command it needs is automatically created in the saved Terminal Commands list. Prompt Library MCP lets the agent reuse your saved prompts when it judges a prompt is legitimate, and ask whether a long, well-structured prompt deserves to be saved. Browser MCP lets the agent navigate, click, type and screenshot the embedded browser to test what it just shipped.
The point of AgentsRoom MCP is choice. Every action you do by hand in the AgentsRoom IDE has a manual path and an MCP path. You stay in control : you can keep things manual, you can automate everything, you can mix the two on the same project. AgentsRoom MCP is what lets the IDE accumulate value over time instead of forcing you into a single workflow.
Four MCP servers, one agent-pilotable IDE
Each AgentsRoom MCP server gives your AI agents CRUD-style tools on a real part of the IDE. Read what's there, create what's missing, update what changed, run what needs to run.
Backlog MCP
CRUD on the project backlog. The agent reads tickets, creates new ones, updates status, links agents to tasks. A ticket can be named by its short number (#42) as well as by its id, and every result carries that number in a ref field. Wired to the public backlog, AgentsRoom MCP also lets the agent reply directly to user feedback : confirm a ticket has been fixed, ask for clarification, mark a feature request as planned. Your AI agent finally talks back to your users without you typing the message yourself.
Terminal Commands MCP
CRUD and run on saved dev commands. The agent can list every command saved in the project, start the right one (yarn dev, docker compose up, pnpm api:dev) and watch it run in the AgentsRoom Dev Terminals panel. If a command does not exist yet, the agent creates it and the new entry shows up live in the Terminal Commands modal, ready to be reused manually next time. The same server also exposes your saved SSH connections: ssh_list shows the remote machines usable in this project, ssh_exec runs a command on one of them and hands back its output, ssh_transfer moves one file in or out, and ssh_connect opens one as a terminal tab you can watch. Credentials never reach the agent, the desktop injects them itself. And when the agent's own sandbox refuses a command, commands_propose puts it in front of you as a Play / Reject card in that agent's pane instead of printing it for you to run somewhere else: hit Play and it runs in a real terminal tab the agent then reads itself.
Prompt Library MCP
CRUD on the project Prompt Library. The agent can list saved prompts and reuse one when it judges the use case is legitimate. The agent can also propose to save a long, structured prompt to the library : when you write a serious brief, AgentsRoom MCP asks whether you want to keep it for next time. Your prompt library grows naturally instead of staying empty.
Browser MCP
The AgentsRoom Browser MCP server (agentsroom-browser) lets the agent drive the embedded Chromium browser : navigate, click, type, screenshot, evaluate JavaScript, read console logs. A QA Engineer agent can open the localhost site, exercise the feature it just shipped, and verify the result before claiming the work is done.
Your servers, your databases and your secrets, as agent tools
AgentsRoom is where you keep your SSH connections, your database connections and your credentials. AgentsRoom MCP is how an agent uses them without ever holding one. The agent works with references, the app works with the secret.
SSH connections
The agent lists the servers saved for the current project and for your account, opens one as a visible terminal tab, and can propose a machine you have not saved yet. Passwords, passphrases and private keys stay between you and your OS keychain : the desktop injects the credential itself, at the moment the terminal opens.
Databases
The agent lists your saved MySQL, PostgreSQL and MongoDB connections, walks the databases, tables and columns without writing a query, runs one statement at a time and gets a capped result set back: reads only, unless you made that connection writable and turned on Allow agents to write. It can also propose a new connection, including the SSH or AWS SSM tunnel the database should be reached through.
Secrets
The agent lists the names of the secrets in your vault and references one as {{secret:NAME}} in a command or in an environment. AgentsRoom resolves it at launch, inside the main process, into the environment of the process being started. There is no tool that returns a value.
What an agent can do
- List the SSH connections, the database connections and the secret names available in the current project and on your account.
- Open a saved SSH connection as a real terminal tab you can read, type into and close.
- Explore a database schema, its tables and its columns without writing a line of SQL.
- Run one statement per call and get a capped result set back: a read, unless you made the connection writable and turned on Allow agents to write for it.
- Reference a secret by name so the value lands in the process it is about to start, and nowhere else.
- Propose a new SSH, AWS SSM or database connection pre-filled with what it found in your project, and unless you turn off "Agents can manage connections" (Settings, on by default), save, update and delete connections outright too, the credential being copied from a vault secret the agent only names.
What an agent cannot do, enforced in code rather than promised in a prompt
- Receive a password, a passphrase, a private key or a secret value. No tool returns one, because an instruction that only lives in a prompt is an instruction a long conversation eventually loses.
- Write to a database you did not open to it. db_query refuses anything but a read unless the connection is writable and you turned on Allow agents to write for that connection (off by default). The check lives in the desktop app rather than in the MCP process, because you do not need a password to run a DROP.
- Save a connection once you turned that off. "Agents can manage connections" (Settings) cuts the direct-save tools back to proposals that only pre-fill the creation form, which you review and complete; and even while it is on, a credential value is never accepted, only the name of a vault secret.
- Work out of sight. Every session it opens is a visible tab in the app, and every connection is scoped to one project or to your account.
Your agents write to each other, through the same server
The same AgentsRoom MCP server turns the saved agents of a project into permanent members with a real address and a real inbox. Seven tools carry it, so a Claude Code agent and a Codex agent in the same project exchange messages without either of them knowing what the other runs on.
- A message is written to disk before delivery is attempted, so an agent that is offline, a CLI that crashes or an app restart never loses one.
- Delivery is held while the recipient is thinking, and held while it is waiting on an answer from you, because writing into that prompt would answer in your place.
- Read, accepted, declined and replied are separate receipts, and reading the body is what marks a message read.
- The sender is stamped by the server from the calling CLI's own identity, so an agent cannot send under another agent's name.
AgentsRoom MCP : four MCP servers (Backlog, Terminal Commands, Prompt Library, Browser) wire the AgentsRoom IDE into your AI coding agents.
Why AgentsRoom MCP is the unique edge
Most IDEs treat AI coding agents as guests. The agent gets a chat box and a few tool calls, but the IDE itself is a black box. You still drag the ticket. You still click the play button on the dev server. You still copy the prompt by hand. The agent watches you do the IDE work it could do for itself.
AgentsRoom MCP flips the model. The IDE is exposed through MCP, the same protocol the agent already speaks. Backlog, Terminal Commands, Prompt Library and Browser become tool calls. The agent moves your ticket to In Progress, starts the dev server, picks the right prompt, opens the browser and tests : the same buttons you click, exposed as agent tools.
This matters because every action becomes optional. You still click the play button when you want. You still drag the ticket when you want. But when an agent is mid-loop and would benefit from doing it itself, AgentsRoom MCP gives it the path. No more shuttling between the agent and the UI. No more screenshots pasted into chat to describe what you just did. The IDE and the agent share the same surface.
AgentsRoom MCP is also what makes AgentsRoom defensible as a tool. Anyone can ship a chat box. Very few products ship an IDE that exposes its own state to the agent through MCP : a Backlog MCP that survives compaction, a Terminal Commands MCP that creates missing entries, a Prompt Library MCP that asks before saving, a Browser MCP wired to the embedded browser. AgentsRoom MCP is the part of AgentsRoom that compounds : the more the IDE accumulates, the more your agents can do.
Manual when you want it. Automated when you don't.
AgentsRoom MCP never replaces the manual UI. Every Backlog MCP action, every Terminal Commands MCP run, every Prompt Library MCP save and every Browser MCP click has a fully manual equivalent in the AgentsRoom IDE. You drag the ticket, you click the play button, you copy the prompt, you open the browser tab. Nothing is forced.
What changes is the option. With AgentsRoom MCP wired in, an agent in the middle of a long task can also do these things itself, without breaking out of its loop. It can mark the ticket fixed, restart the dev server it just changed, save the brief you wrote as a reusable prompt, open the browser to verify the fix. Each of those is one MCP tool call away.
The result is leverage without autopilot. You decide which actions stay manual, which actions you want the agent to handle, and AgentsRoom MCP exposes the right tools for both. Power users automate everything ; new users keep things manual ; both run on the same IDE.
An MCP-driven IDE, not a chat box on top of a terminal
Most coding agents see the world through stdin, stdout and the filesystem. AgentsRoom MCP gives them more.
AgentsRoom MCP exposes high-level workspace concepts as MCP tools. The agent does not have to grep your project to know if a backlog ticket exists, it calls backlog_list. It does not have to guess your dev command, it calls commands_list. It does not have to invent a prompt template, it calls prompts_list. It does not have to spawn a Playwright config, it calls browser_navigate. AgentsRoom MCP is the layer that makes the IDE legible to the agent.
On the desktop, the AgentsRoom MCP servers run as small local processes. The bridge between an MCP subprocess and the AgentsRoom desktop runs on 127.0.0.1 only, on an OS-assigned port, authenticated with a 32-byte hex token regenerated at every boot. Your data never leaves the machine. AgentsRoom MCP is local-first by design : the IDE and the agent talk to each other on localhost, not through somebody else's cloud.
How AgentsRoom MCP works in practice
Open a project in AgentsRoom
AgentsRoom MCP servers are auto-registered with every agent in the project. Claude, Codex, GitHub Copilot CLI, Cursor and 13 other agent CLIs all see the same MCP tool surface : Backlog MCP, Terminal Commands MCP, Prompt Library MCP and (when enabled) Browser MCP. A small pill in each agent's header shows whether its CLI actually loaded them. In Settings > AgentsRoom tools you choose which tool families your agents get, for every project, a folder, one project or one agent: a family you switch off is not even listed to the agent.
Send a brief to your agent
Drop a real task on the agent : 'Add Stripe checkout to the demo, add a dev server command for the demo if it does not exist, then test in the browser and reply to the public backlog ticket'. The agent treats AgentsRoom MCP as part of its tool belt and starts pulling the right levers.
The agent updates the backlog
Through Backlog MCP the agent moves your ticket to In Progress, posts an update, and at the end of the run flips the ticket to Done with a summary. If the ticket came from the public backlog, AgentsRoom MCP lets the agent reply to the original reporter with the resolution.
The agent runs the dev server
Through Terminal Commands MCP the agent calls commands_list, picks the existing 'Stellar Demo Dev' command and starts it with commands_run. The Dev Terminals panel shows the server running, exactly as if you had clicked play yourself. If the command was missing, the agent uses commands_create first and the new tab appears in the Terminal Commands modal.
The agent reuses or saves a prompt
Through Prompt Library MCP the agent fetches a relevant saved prompt with prompts_get and applies it to the current task. When you wrote the brief manually and it looks reusable, AgentsRoom MCP asks whether it should save your brief as a new entry in the Prompt Library.
The agent tests in the browser
Through the Browser MCP server (agentsroom-browser) the agent opens the localhost preview, exercises the new flow, takes a screenshot, reads console logs and confirms the feature works before reporting back. End-to-end verification, end-to-end logged inside the same AgentsRoom IDE.
Why AgentsRoom MCP makes AgentsRoom more complete
Six things AgentsRoom MCP gives you that a regular IDE plus a chat box cannot.
An agent-driven IDE
The agent does not just write code. With AgentsRoom MCP it also drives the backlog, the dev terminals, the prompt library and the browser. The IDE itself becomes part of its workflow.
Manual or automated, never forced
Every AgentsRoom MCP action keeps a manual path. You decide what stays click-driven, what becomes agent-driven, and the IDE supports both.
Local-first, loopback only
AgentsRoom MCP servers talk to the IDE over 127.0.0.1, on an OS-assigned port, with a per-boot auth token. Your data never leaves your machine.
Provider-agnostic
AgentsRoom MCP works with Claude Code, Codex CLI, GitHub Copilot CLI, Cursor and 13 other agent CLIs. Any MCP-aware coding agent gets the same Backlog, Terminal Commands, Prompt Library and Browser tools.
Wired into Agent Teams
When you orchestrate a multi-agent workflow, every node inherits the AgentsRoom MCP tool surface. Dev hands off to QA, QA opens the Browser MCP to validate, both update the Backlog MCP.
Low overhead, high leverage
AgentsRoom MCP servers run locally: one background process serves all your agents, about 1 MB each, and adds zero round-trip latency between the agent and the IDE.
Drop the brief once, AgentsRoom MCP does the routing
Here is the kind of prompt that AgentsRoom MCP unlocks. Notice how each verb maps to one MCP server : the agent reads the backlog, runs the dev command, reuses a saved prompt, opens the browser and reports back. Without AgentsRoom MCP you would do all of this by hand.
Example prompt
Pick the top public backlog ticket through Backlog MCP, mark it In Progress, ship the fix, run the demo dev server through Terminal Commands MCP (create the command if it is missing), use the saved 'commit message generator' prompt from Prompt Library MCP, open the Browser MCP to confirm the fix on the live preview, then update the ticket to Done and reply to the user.
FAQ
What is AgentsRoom MCP exactly ?
AgentsRoom MCP is a set of MCP servers that expose the AgentsRoom IDE to the AI coding agents running inside it. Today AgentsRoom MCP ships four servers : Backlog MCP, Terminal Commands MCP, Prompt Library MCP and Browser MCP. Each server gives the agent CRUD-style tools on a real part of the IDE.
Which AI coding providers can use AgentsRoom MCP ?
AgentsRoom MCP follows the standard Model Context Protocol. All 17 agent CLIs AgentsRoom supports (Claude, Codex, GitHub Copilot CLI, Cursor and 13 more) support MCP and can use AgentsRoom MCP out of the box. Any other MCP-aware agent works too.
Can my agent reply to public backlog tickets through Backlog MCP ?
Yes. Backlog MCP exposes backlog_create and backlog_update so the agent can post status updates and replies on a public backlog ticket. When a user files a bug from the public backlog, AgentsRoom MCP lets the agent confirm the fix, ask for clarification or mark the request as planned without you typing the message yourself.
How does Terminal Commands MCP create commands ?
When the agent calls commands_run with a command that does not yet exist, AgentsRoom MCP creates the entry through commands_create and stores it in the project's saved commands. The new tab appears live in the Dev Terminals panel and you can reuse it manually next time. Terminal Commands MCP teaches the IDE the commands the agent finds, so the project compounds value.
Does Prompt Library MCP modify my prompts without asking ?
No. Prompt Library MCP is read-by-default. prompts_list and prompts_get let the agent reuse a saved prompt when it judges the use case is legitimate. prompts_save is proposed by the agent when you write a long, structured brief that looks reusable, but the actual save is your call. AgentsRoom MCP keeps you in the loop on writes.
How is AgentsRoom MCP different from a generic MCP server ?
Generic MCP servers expose external systems (databases, APIs, files, browsers) to an agent. AgentsRoom MCP exposes the IDE itself : the backlog, the dev terminals, the prompt library, the embedded browser. It is the IDE talking to the agent, not just the world around it. That is the unique edge.
Is AgentsRoom MCP local-first ?
Yes. AgentsRoom MCP servers run as small local processes on your own machine. The bridge to the AgentsRoom desktop runs on 127.0.0.1 only, on an OS-assigned port, with a 32-byte hex auth token regenerated at every boot. AgentsRoom MCP never routes your data through a third-party cloud.
Does AgentsRoom MCP replace manual actions ?
No. AgentsRoom MCP adds an automated path on top of the manual UI. Every backlog drag, every Dev Terminals play button, every Prompt Library save and every Browser click is still available manually. AgentsRoom MCP is the option to let an agent do the same thing for you when it makes sense.
Can I use AgentsRoom MCP from inside Agent Teams ?
Yes. Agent Teams nodes inherit the AgentsRoom MCP tool surface automatically. A Dev node can run a dev command, a QA node can open the Browser MCP and verify the result, both can read and write the Backlog MCP. AgentsRoom MCP is what lets multi-agent workflows actually touch the workspace.
Where can I see AgentsRoom MCP in action ?
The video on this page shows the Terminal Commands MCP server in action : the agent runs a dev command and the new terminal tab shows up in the Dev Terminals modal with the saved commands. The same pattern applies to Backlog MCP, Prompt Library MCP and Browser MCP. The result is a full agent-pilotable IDE built on AgentsRoom MCP.
Can my agents use my SSH connections, my databases and my secrets ?
Yes, through the same AgentsRoom MCP server. ssh_list lists the servers saved for this project, ssh_exec runs a command on one of them and returns its output with the exit code, ssh_transfer moves one file to or from a server without the file ever going through the agent conversation, and ssh_connect opens one as a visible terminal tab. db_list and db_schema list your MySQL, PostgreSQL and MongoDB connections and explore a schema. db_query runs one statement with a capped result set; writes require a writable connection and your explicit Allow agents to write setting. secrets_list returns the names of your secrets so a command or an agent environment can reference one. And ssh_connection_new and db_connection_new let an agent propose a connection you have not saved yet.
Do those tools ever hand a credential to the agent ?
No tool returns a password, a passphrase, a private key or a secret value: the desktop injects the credential itself when it opens the terminal, database socket or process. Database writes are a separate permission: they run only when you explicitly enable them on a writable connection. A proposed connection only pre-fills a form that you review, complete and save yourself.
Can my agents send messages to each other through AgentsRoom MCP ?
Yes. Seven tools on the same server (agents_list_live, agents_send, agents_message_status, agents_read_inbox, agents_reply, agents_ack, agents_report_status) let the saved agents of a project write to each other as permanent members, whichever CLI each one runs. Messages are stored before they are delivered, so an offline agent or a restart never loses one, and the sender is stamped by the server rather than passed as an argument.
Can my agents read my usage limits and manage my quota rules?
Yes. usage_overview returns every quota bar of the Usage panel for each signed-in account, with its reset time and, on weekly and monthly bars, the pace: what may be spent by now against what is spent. An agent can check it before starting a long task, or send the work to the provider with the most room left. quota_rules_list, quota_rules_save and quota_rules_delete read, create, edit, reorder and remove your quota rules one at a time. The order of the list is the priority, and a rule written by an agent shows up in the Quota management card like one you wrote yourself. The tools tell the agent to check with you before it adds a rule that blocks or redirects launches, and before it deletes one.
Can I switch off the AgentsRoom MCP tools I do not use?
Yes. Settings > AgentsRoom tools lists every tool family (backlog, project memory, Agent Mail, agent consoles, teams, SSH, databases, QA Bot and more) with its size in the agent's context. Switch off a family, or a single tool, for every project, then override it per folder, per project or per agent. A switched-off tool is not even listed to the agent, its instructions leave the startup prompt and a call to it is refused, so an agent that runs on your own memory or backlog system never gets contradictory instructions. You can also rewrite the description of any tool, and the app tells you when an update changes the original text.
Can my agents hand work to each other and follow it up through AgentsRoom MCP?
Yes. Four tools on the same server (assignments_create, assignments_list, assignments_mine, assignment_update) put a delegation on record: the brief and its acceptance criteria travel as a message between agents, the agent that receives it reports its progress and submits a structured result, and the agent that delegated verifies it or returns it. assignments_list gives each delegation's next step, decided by fixed rules from its state, its check-in delay and the live console of the agent doing the work, so a coordinator acts on a fact instead of rereading its mail. Your own scripts read the same data from the app's local read-only API.
Goes well with
Agent Messaging
The seven tools above, explained in full: a permanent roster, a persistent inbox, an offline queue and real receipts, across every CLI.
Agent Delegation
The dev to QA agent delegation flavor : your dev agent calls run_qa_test through MCP, an ephemeral QA agent runs the test on a cheaper model and reports back.
Dev Terminals
The terminal manager Terminal Commands MCP drives. Saved commands per project, detachable window, AI command generation, live status.
Browser Automation
The embedded browser the Browser MCP drives. Real Chromium per project, persistent session, end-to-end test loop for AI agents.
Backlog Task Board
The kanban board the Backlog MCP drives. Drag a task to In Progress, an agent picks it up. Wire it to your public backlog for client-driven work.
Prompt Library
The prompt manager the Prompt Library MCP drives. Save reusable prompts per project, share with your team via git, send to any agent in one click.
Agent Teams
Multi-agent workflow editor where every node inherits the AgentsRoom MCP surface. Dev to QA hand-off, conditional edges, feedback loops.
MCP Dashboard
Plain-English explainer of the Model Context Protocol and how AgentsRoom positions itself as the visual command center for MCP-powered Claude Code agents.
Further reading
AgentsRoom Now Supports Grok Build
Grok Build, xAI's terminal coding agent, is now a first-class provider in AgentsRoom. Run it next to Claude, Codex, Antigravity CLI, OpenCode and Aider, and switch mid-conversation.
AgentsRoom Now Supports Kimi Code
Kimi Code, Moonshot AI's terminal coding agent, is now a first-class provider in AgentsRoom. Run it next to Claude, Codex, Antigravity CLI, OpenCode, Aider, Grok Build and Mistral Vibe, and switch mid-conversation.
AgentsRoom Now Supports Mistral Vibe: A Sovereign European Coding Agent
Mistral Vibe, the terminal coding agent from Paris-based Mistral AI, is now a first-class provider in AgentsRoom. Run it next to Claude, Codex, Antigravity CLI, Grok and Aider, and build an end-to-end European AI dev stack.
Pilot the IDE, or let your agents pilot it
Download AgentsRoom and try AgentsRoom MCP. Backlog MCP, Terminal Commands MCP, Prompt Library MCP and Browser MCP, all wired to Claude, Codex, GitHub Copilot CLI, Cursor and 13 other agent CLIs, on your machine, free.
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.