ExecWarden Docs

Overview

ExecWarden sits between agents and the tools they use.

Give agents scoped per-tool access and approval points for risky actions without placing broad provider credentials in the agent environment. ExecWarden records the calls it mediates.

What ExecWarden Does

Choose whether the agent loop runs in Codex CLI, Claude Code, OpenCode, another compatible MCP client, or a hosted ExecWarden session. ExecWarden grants access per tool, asks before risky actions, and records the capability calls it mediates.

Choose the agent loop

Use an ExecWarden-hosted session, or keep the loop in Codex CLI, Claude Code, OpenCode, or another compatible MCP client.

Choose the execution shape

Start without a machine for lightweight work, or connect customer-configured capacity for machine-shaped work. The hosted-agent and execution-capacity pages describe the current beta paths.

Scope each tool grant

Expose the repo boundary, API action, package tool, or imported MCP tool instead of copying broad provider credentials into the agent environment.

Pause risky actions

Put comments, PR actions, deploys, credits, account changes, and other writes in Ask mode so the exact payload pauses before execution.

Inspect the record

Use the activity log, transcripts, approval decisions, provider results, identity status, and retry failures to see what happened.

Agent identity with a GitHub pull request capability scoped to a repository and branch pattern
GitHub example: the identity detail makes one per-tool boundary concrete through its capability, mode, selected repository, and branch pattern.

Three Immediate Paths

Connect an MCP client

Use the tested Codex CLI, Claude Code, or OpenCode path when you want an existing client to call scoped ExecWarden tools.

Run a hosted agent

Use lightweight gateway-backed work without a machine, or connect customer compute for a machine-backed interactive session.

Completed Run

A verified ExecWarden-hosted agent with connected execution capacity handled a seeded bug, a selected repo and agent/* branch boundary, a failing test, a patch, two passing tests, a PR, and no merge. See the completed run summary for the full sequence.

Seeded bugA known issue was seeded before the run.
Branch boundaryThe run allowed one selected demo repository and agent/* branches.
Run resultConnected compute supplied the repository and tests; the hosted agent completed the patch and opened a PR through governed GitHub access.
Authority shapeThe unattended run used scoped Allow grants.

Choose Another Task

GitHub quickstartUse the GitHub Quickstart when code is the first system you want to govern.
MCP client setupUse the MCP Quickstart to connect Codex CLI, Claude Code, or OpenCode.
Task selectionUse Use Cases to choose repo work, internal API work, support work, or event-driven work.
Current supportUse Current Support to see what is usable now and what still needs polish.