Access Sparse Agent from another Agent: CLI and MCP guide

In Codex, Claude Code or WorkBuddy, you can send a follow-up to an existing Sparse Agent and check progress from the same client. CLI and MCP connections use that Agent's existing goal, memory and conversation. You can also view the same chat history on the website.

Availability: enabled (October 13, 2026). Sign in to the web workspace, then open Settings → CLI / MCP to create a credential, copy installation commands or configure MCP.

Choose a connection method

The CLI runs on your computer; its Skill tells your client how to call it. MCP exposes Sparse Agent as tools without installing the Sparse Agent CLI. Your MCP client must support Streamable HTTP, and browser authorization also requires OAuth support.

Prepare an existing Agent

Follow the main guide to sign in, configure your model and cloud environment, create an Agent and set its Goal. Confirm its environment is ready, then decide which Agents to authorize.

CLI users open Settings → CLI / MCP, select an Agent, name the connection, choose an expiry and create a credential. It appears only once; use it yourself in the terminal's hidden prompt. MCP users can start browser authorization directly from their client and select Agents on the consent page.

Clients can access only authorized Agents. Your model key and cloud provider token remain in their settings. Keep client credentials out of Agent chat, command arguments and project files.

Install the CLI and sign in

On macOS or Linux, run:

curl -fsSL https://sparse-agent.com/install.sh | sh -s -- --client all

To configure one client, replace all with codex, claude or workbuddy. The CLI supports Intel and ARM64 on macOS / Linux, plus Windows x64. The Downloads page also provides packages for each supported system.

On Windows, run in PowerShell:

Invoke-WebRequest https://sparse-agent.com/install.ps1 -OutFile install-sparse-agent.ps1
.\install-sparse-agent.ps1 -Client all

Open a new terminal, check the tool and sign in:

sparse-agent --version
sparse-agent --help
sparse-agent login
sparse-agent agents --json

login asks you to paste the credential without displaying your input. Then agents lists authorized Agents; keep the name and ID of the one you intend to use. If the command cannot be found, check PATH for the default directory: ~/.local/bin on macOS / Linux, or %LOCALAPPDATA%\SparseAgent\bin on Windows.

Let your client discover the Skill

Installation adds the sparse-agent Skill to ~/.agents/skills/sparse-agent/ for Codex and ~/.claude/skills/sparse-agent/ for Claude Code. Start a new client session and ask it to read sparse-agent --skill.

For WorkBuddy, the installer exports a Skill ZIP and prints its location. Choose Add Skill → Upload Skill, import it, enable it and start a new task. Downloads also provides workbuddy-sparse-agent.zip. See WorkBuddy's official Skill instructions for the interface steps.

Installing the Skill, enabling it and signing in are separate steps. If your client cannot run local commands, use a supported MCP connection.

Send a follow-up from your client

For an existing Agent named “Content operations assistant”, try:

Read sparse-agent --skill and list my authorized Agents.
Find “Content operations assistant” and send one Run Now request: summarize today's progress, completed work, problems and next steps.
Keep reading its public intermediate messages and summarize the final reply for this request.
If the name is ambiguous, ask me to choose. Keep the request ID and cursor to resume reading after an interruption.

Run Now sends a follow-up to an existing Agent. If it is busy, the request may queue. A QR-code or sign-in wait still requires you to complete Needs your help on the website. Sending a message does not automatically change the long-term Goal.

You can call the CLI directly. Replace AGENT_ID with the ID from the list command:

sparse-agent run-now --agent AGENT_ID --message "Summarize today's progress and keep me updated as you work" --follow --jsonl

--follow waits for public progress and the final reply; --jsonl lets your client read one JSON record at a time. Progress consists of messages the Agent explicitly shares. Completion means this follow-up has been handled; the long-term goal continues on its existing schedule.

Connect through MCP

Add this endpoint in your client's MCP settings and select HTTP / Streamable HTTP:

https://sparse-agent.com/mcp

For Codex, configure it from a terminal:

codex mcp add sparse-agent --url https://sparse-agent.com/mcp
codex mcp login sparse-agent

For Claude Code, run:

claude mcp add --transport http sparse-agent https://sparse-agent.com/mcp

Then enter a Claude Code conversation, use /mcp, select the server and complete browser authorization. See the official Claude Code MCP documentation for options. Interface labels may vary between versions.

Sign in to Sparse Agent in the browser, verify the service address and select the Agents to authorize. The client can then use four tools:

Tell your client: “List my authorized Agents and ask the content operations assistant for today's progress. Submit once, keep the request ID and read until the request completes or fails.” Generate a unique request_id for each new instruction, and retain it with the returned cursor for later reads.

Some clients display live progress; others need to keep reading events. Each tool call can wait up to 30 seconds; continue reading after it returns. OAuth clients with refresh-token support can renew automatically. Access tokens last up to one hour; authorize again after 30 days without renewal or 90 days from initial authorization. A client with only manual Bearer configuration can use a website-generated credential, renewed according to the expiry shown there.

Resume after waiting ends

The CLI waits up to 300 seconds by default; use --timeout to adjust it. A timeout (exit code 3), Ctrl-C or a closed terminal ends local waiting only. A request already accepted in the cloud continues.

Keep the output's REQUEST_ID and latest CURSOR to read the same request:

sparse-agent events --agent AGENT_ID --request-id REQUEST_ID --after CURSOR --follow --jsonl
sparse-agent request --agent AGENT_ID --request-id REQUEST_ID --json

If the network fails during submission, query the printed request ID first. If a retry is needed, use the same --request-id and original message to avoid duplicates. Use a new ID for a new instruction. When a cursor expires, the CLI recovers current records and reports that earlier progress may be missing.

Revoke access and resolve problems

Revoke authorization in Settings → CLI / MCP to stop API and MCP access and further renewal. sparse-agent logout removes this computer's saved credential; revoke it on the website to stop other copies.

For your first connection, authorize one Agent, ask for progress and compare the message and reply with the website. Add other Agents once you have confirmed the workflow.