Agent Plugin MCP Server

In Parallel

Bring your team's context into conversations with your AI assistant. Catch up on priorities, find relevant decisions, and share useful findings with your team.

Get started

You'll need an In Parallel account with access to your team's workspace.

Claude Code, Codex, or Cursor

Run the setup command. It asks which assistants to set up, then installs for your user:

npx -y github:in-parallel-oy/agent-plugin

Needs Node 20.12+. Claude Code and Codex also need their own CLI on your PATH. Cursor users should reload Cursor afterwards.

Later on, the same command takes doctor to check the installation and uninstall to remove it. Add --help for everything else.

Claude Desktop or ChatGPT

Open the app's plugin browser and install In Parallel from your team's plugin library. If it isn't listed, add https://github.com/in-parallel-oy/agent-plugin as a plugin source, or ask your administrator to make it available.

Claude Projects

Threads in a Claude project reach In Parallel through your claude.ai connector; the project conversation itself can't use In Parallel, only its threads can.

1. Connect In Parallel under claude.ai → Customize → Connectors.
2. Get the claude-projects ZIP from your administrator, or build it from a checkout of this repository with node scripts/package.js --client claude-projects. It holds the skill only, with no MCP server, hooks or scripts.
3. Add it in Project settings → Plugins.

If your workspace offers it, Copy instructions for a Claude project on the workspace's What's set up page gives you project instructions to paste in.

Then

Sign in to In Parallel, choose the workspaces you want to connect, and start a new conversation with your assistant.

Try it

• “What needs my attention this week?”
• “Catch me up on [project].”
• “Record this finding for my team: …”

What the plugin does on your machine

The plugin has two parts: the in-parallel skill and a connection to the In Parallel server. It runs no hooks or scripts on your machine and keeps no local files. The files under experimental/ and the hook scripts in scripts/ are not part of it; see Experimental features.

What is sent to In Parallel

Only what your assistant sends through the In Parallel connection (https://www.in-parallel.ai/mcp unless you chose another environment during setup). The skill asks your assistant to:

• read team context, such as briefings, meetings, decisions and wiki pages
• record findings you want to share with your team

Sign-in

You sign in through your assistant app's connection sign-in, which opens In Parallel in your browser. The app keeps the sign-in. The plugin's scripts never see, read or store a password or token.

The setup command

npx -y github:in-parallel-oy/agent-plugin downloads this repository and its two setup dependencies. It copies the plugin to ~/.in-parallel/setup/ for Claude Code and Codex, or ~/.cursor/plugins/local/in-parallel for Cursor, and registers it with the claude or codex command. It writes your chosen server URL into that copy. For the Development environment or a custom server, it also adds the experimental features.

Before it installs, it reads your existing MCP settings to find a conflicting In Parallel entry: ~/.claude.json (or the one in CLAUDE_CONFIG_DIR), .mcp.json in the current project, and ~/.cursor/mcp.json and .cursor/mcp.json. It never changes those files.

Uninstall

If you used the setup command, run:

npx -y github:in-parallel-oy/agent-plugin uninstall

If you installed from a plugin browser, remove In Parallel there. If you used the experimental work journal, also delete ~/.in-parallel to remove its local cache.

Experimental features

Features the In Parallel server has not yet enabled for every team live in experimental/. Plugin browsers never install them. The setup command adds them only when you choose the Development environment or a custom server, never for Production.

Work journal (experimental/work-claims)

Adds the in-parallel-work skill and two hook scripts. The skill asks your assistant to report the tasks it works on without being asked each time: a short title, a one to three sentence description, progress notes, and links such as the repository or pull request URL.

In Claude Code and Codex, scripts/context.js runs when a session starts, when you send a message, and when a subagent starts. scripts/remember-claim.js runs after your assistant calls the In Parallel announce_work, get_work or link_work_subject tools.

In Cursor, scripts/context.js runs when a session starts and after shell, file write, file delete, task and MCP tool calls. scripts/remember-claim.js runs after each MCP tool call and ignores every server except In Parallel.

Each hook has a 10-second limit and never blocks your session. Neither script connects to In Parallel.

context.js:

• reads the checkout's origin remote URL and current branch with git. Any user name or password in the remote URL is dropped.
• asks GitHub for a pull request on the current branch, if the GitHub CLI gh is installed. gh uses your existing GitHub sign-in.
• reads the local cache described below, and the MCP URL from the plugin's own .mcp.json or mcp.json.
• adds one line of context to the conversation: the repository URL, branch, pull request URL, In Parallel server URL, a session ID, a workspace hint, and the titles and status of this session's open work entries.
• writes a small marker in ~/.in-parallel/sessions/ so the line is only repeated when something changes, and notes the time in the local cache. Markers older than seven days are deleted.

remember-claim.js reads the In Parallel tool's reply and saves a handle for your own work entries in ~/.in-parallel/contributions.json: the entry ID, version, status, title, latest update note, workspace ID, linked outcome ID, start time and the folder you were working in. Finished or cancelled entries keep only their ID, version, status and server. It also records when context, reads and reports last happened, so doctor can check the setup.

The folder is readable only by you. The scripts send none of it anywhere; the titles and workspace hint appear in the context line above.

Need help?

Ask your assistant: “Check my In Parallel connection.”

If you can't install the plugin or access your workspace, ask your workspace administrator for help.

Publishing to the MCP registry

server.json lists the hosted In Parallel MCP server in the official MCP registry: the context layer for your company's AI, which gives Claude, ChatGPT, Copilot and other MCP clients your team's decisions, meetings and plans, with sources and scoped to each user's permissions. PulseMCP and Glama ingest new entries from the official registry; Smithery needs its own submission at smithery.ai.

Install mcp-publisher (brew install mcp-publisher) and check the file:

mcp-publisher validate

The name ai.in-parallel/in-parallel needs proof that you control in-parallel.ai. Both proofs use an Ed25519 key, which needs OpenSSL 3 (brew install openssl@3; macOS LibreSSL lacks Ed25519):

OPENSSL=/opt/homebrew/opt/openssl@3/bin/openssl
$OPENSSL genpkey -algorithm Ed25519 -out key.pem
PUBLIC_KEY="$($OPENSSL pkey -in key.pem -pubout -outform DER | tail -c 32 | base64)"
PRIVATE_KEY="$($OPENSSL pkey -in key.pem -noout -text | grep -A3 'priv:' | tail -n +2 | tr -d ' :\n')"

Then pick one:

• DNS. Add a TXT record on the in-parallel.ai apex (not a subdomain) with the value v=MCPv1; k=ed25519; p=${PUBLIC_KEY}, wait for dig +short TXT in-parallel.ai to show it, then run mcp-publisher login dns --domain in-parallel.ai --private-key "$PRIVATE_KEY".
• HTTP. Serve the single line v=MCPv1; k=ed25519; p=${PUBLIC_KEY} as text/plain at https://in-parallel.ai/.well-known/mcp-registry-auth. The registry fetches the apex and does not follow redirects, so the apex-to-www redirect must skip this path. Then run mcp-publisher login http --domain in-parallel.ai --private-key "$PRIVATE_KEY".

After logging in, publish:

mcp-publisher publish

Without domain access, rename the server to io.github.in-parallel-oy/in-parallel and log in with mcp-publisher login github as an Owner of the in-parallel-oy GitHub organization.

Keep key.pem out of the repo. Each published version is permanent, so bump version in server.json together with plugin.json and publish again when the listing should change.

License

Apache-2.0.