Agent loop with arg - arg.ai

When the local workflow includes a model — Claude Code, Codex, a custom agent — you typically don’t want it touching the filesystem just to ferry data. Wire arg.ai into the session as an MCP server and the agent reads and writes the workspace directly, with the file history acting as your audit log.

# Workspace-scoped — exposes one workspace as MCP tools
claude mcp add -t http arg-workspace https://api.arg.ai/mcp/{workspace_id} \
  -H "x-api-key: $ARG_API_KEY"

The agent now has read_file, write_file, edit_file, grep, run_bash, browse_url, screenshot_url, extract_webpage_data, upload_file, and more — see the MCP guide.

What you layer on top

Anything you already give Claude Code locally, the agent can combine with arg.ai writes:

Local capability	What the agent does	Where it lands in arg
Playwright / Chrome DevTools MCP	Drive your real Chrome to scrape gated pages, fill forms, take screenshots	`/research/...`, `/crm/...`
Filesystem access	Read a local repo, summarize, refactor	`/eng/reviews/...`
A database driver	Query Postgres, summarize trends	`/analytics/...`
A local LLM (Ollama, llama.cpp)	Run private inference, post results	`/ml/runs/...`
Any other MCP server (Linear, GitHub, Slack)	Pull tickets/PRs/messages	`/inbox/...`

The pattern is always: local tools provide ingestion and side-effects; arg holds the durable state. When tomorrow’s session runs, it read_files yesterday’s output and continues — no scratch files, no ~/Downloads/leads-final-v3.csv.

Why split it this way

Side-effects stay local. Logged-in browser sessions, DB credentials, on-device models — these belong on your laptop, not in the cloud.
State stays durable. Anything worth keeping ends up at a stable workspace path, viewable in the UI, queryable over REST, readable by another agent.
Re-entrant. Tomorrow’s run reads yesterday’s output and appends — no duplicate work.
Multi-tenant. Switch workspace_id to point the same workflow at a different team’s data.