ExecWarden Docs

Agent runtime

Choose lightweight hosted work or a machine-backed session.

ExecWarden offers lightweight gateway-backed agent work and machine-backed interactive sessions. Both keep model access, transcripts, governed tools, approvals, and run state together.

A Hosted Session Is Not Automatically a Machine

ExecWarden can host the model loop, transcript, approvals, and run state without attaching general-purpose compute. That lightweight path can run bounded JavaScript, make bounded public HTTP requests where enabled, and invoke granted capabilities. Connect customer-configured execution capacity when the task needs a real filesystem, shell, processes, git, browser, tests, containers, or private-network access.

Current beta distinction: lightweight gateway-backed agent runs can work without a machine. The environment-backed interactive hosted-model path currently requires a running sandbox volume on connected execution capacity.

Model access

Use a model provider connected to the workspace. Model choice stays separate from whether the session has execution capacity.

Lightweight hosted tools

Use bounded JavaScript, bounded public HTTP where enabled, and granted ExecWarden capability calls without attaching a machine.

Connected execution environment

Attach a sandbox volume on a customer-configured VM, Docker host, or local machine so shell, files, browser, git, tests, and processes have a real place to operate.

Transcript and traces

Keep prompts, tool calls, outputs, approval pauses, and run status together.

Capabilities

Attach exactly the integrations, packages, repo access, and approval status tools the task needs.

See Execution Capacity when the task needs a customer-configured VM, Docker environment, or local machine.

Good First Hosted-Agent Tasks

API-backed researchUse the lightweight runtime for bounded public HTTP where enabled, data transformation, and granted API calls without attaching a machine.
Repo investigationAttach customer-configured compute, then ask the agent to inspect repositories and report concrete findings before granting write authority.
Patch and verifyLet an agent edit files and run tests in a sandbox volume, then review the branch push and gate PR creation with approval.
Integration checkGive read-only API capability first, then expand scope once the output is trustworthy.
Workflow draftAsk an agent to propose a workflow definition, manual run shape, access rules, and expected outputs.
Completed hosted runUse Completed Run to see the summarized result of a hosted agent loop with connected compute and scoped GitHub Allow grants for selected repositories and agent/* branches.
Need to expose tools to an external client?Use MCP & API to connect Codex CLI, Claude Code, OpenCode, or another compatible MCP client through an Agent identity.
Need work to start from a request, schedule, or event?Use Workflows to define the graph, step types, pauses, and trigger.