Bring project memory to a coding agent
Give a coding agent useful project decisions through a brief or MCP connection without assuming it has your repository, secrets, or chat history.
dray with
Keep a project moving 3 min read
Give Claude Code a signed-in browser and project memory
Add two MCP servers to the same agent: one for project memory and one for a browser that keeps its logins. The agent looks up what you decided, then checks it on the real site. When a page asks for a password, you sign in yourself and hand the same browser back.
ON THIS PAGE
When one agent needs both
This setup pays off when a task depends on an earlier decision and on a site behind a login. Say you agreed new plan prices last week and want the agent to check the staging checkout against them. Memory holds the decision and where it came from. The browser shows what the site does today.
Plenty of tasks need only one half. If your dev server runs on your own machine and needs no login, a local browser server such as Chrome DevTools MCP is simpler, because it drives Chrome on your computer. If the agent only needs past decisions, connect memory on its own.
Connect each server on its own
Each tool is a separate MCP server with its own credential. dray setup adds Dray’s. In Claude Code, claude mcp add registers the browser server, and the agent then sees the tools from both in the same session. Give each server its own token for this agent, so you can revoke one without touching the other.
Test each server alone before you rely on them together. Ask the memory server for one harmless stored detail and its source. Ask the browser to open a public page and describe it. If a test fails, fix that connection first; two new servers failing at once are hard to tell apart.
Sign in once, then hand the browser back
Tell the agent to stop and ask when it reaches a sign-in page. In a browser like Tallylamp, you open the live view and take control, because watching alone is read-only. Sign in or finish the verification step, then return control, and the agent carries on in the same browser.
The browser keeps that profile, so the next task can start signed in until the site ends the session. Treat control of a signed-in browser as permission to use the accounts inside it. Where you can, sign in to a staging account rather than your own admin account.
A hosted browser runs on a server, so localhost inside it means that server, not your laptop. Tallylamp can open a tunnel to one private host and port on your machine, and it is off by default.
TAKE THIS WITH YOU
A task brief for memory and a browser
An illustrative template. Replace the details with material you are allowed to share.
Task: [what to check or fix] Look up first: [the decision, e.g. the agreed plan prices]. Name the source you used. Site: [staging URL]. If it shows a sign-in page, stop and ask me. Browser permission: [read only / may change staging settings] Never: type a password, touch production, or save page text to memory. Report: what memory says, what the site shows, and any mismatch.
USING THIS WITH DRAY
Keep the context. Check the source.
Dray is the memory half: it brings the decisions from your ChatGPT, Claude and Grok chats into the agent, with their sources. Dray’s agent setup page has the steps for Claude Code and other agents. Claude Code can also save its own sessions to Dray, so what you decide in the terminal is kept with the rest.
Available sources and connections depend on your setup. Read the privacy policy before granting access.
Sources & next reads
Product details are grounded in the guides below. Example projects and prompts are illustrative, not customer results.
- Connect Claude Code, Cursor and other agents to Dray
- Dray help and setup guide
- Claude Code docs: Connect Claude Code to tools via MCP
- Tallylamp: Connect an AI agent to a browser through MCP
- Tallylamp: Sign in for your browser agent with human handoff
- Tallylamp: Set permissions for browser agents
- Tallylamp: Reach localhost from a remote browser