01
Assign the outcome
Choose an Agent context and describe the result in New Work. XpressClaw creates a durable task for the selected harness.
0.2 BETA GPL-3.0 · open source and local-first
The open-source control planefor persistent AI work.
Turn Codex, Claude Code, OpenCode, and 11 more ACP harnesses into durable workers—with retained project environments, task state, schedules, review loops, and a UI you can leave and come back to.
CodexClaude CodeOpenCode+ 11 ACP harnesses
SHIPPED IN THE PUBLIC BETA
A workspace for outcomes, not another terminal.
The current interface is built around Agents, durable tasks, work history, schedules, and workflows—not a terminal stream hidden inside a browser.
Real product UI
THE FIRST AHA MOMENT
From request to reviewed change—without living in terminals.
XpressClaw wraps the native harnesses developers already trust with durable execution state and a control plane that can keep moving after you step away.
02
Implement in a retained environment
Codex—or another harness—works in its configured project container while attempts, progress, and results remain visible.
03
Hand off an independent review
A reusable workflow can pass the draft PR to a separate reviewer context, such as Claude Code, and loop actionable feedback.
04
Resume, redirect, or wait
Inspect the record, add context, cancel work, schedule a follow-up, or let a durable wait resume after external activity.
SHIPPED WORKFLOW
The current beta includes an implementation + review template: one Agent context owns a draft pull request, another reviews it independently, and rejected changes return to implementation before the PR is marked ready.
DURABILITY AROUND THE LOOP
Your coding agents shouldn't be disposable sessions.
The harness keeps its reasoning loop. XpressClaw supplies the environment, execution state, automation, and history around it.
Retained project runtime
Across ordinary turns, an Agent keeps one Docker or Podman container and ACP process instead of rebuilding a disposable sandbox for every prompt.
- Installed SDKs, tools, and caches persist
- Configured startup commands initialize once
- Session resume or load after unavoidable restarts
- Explicit workspace and volume mounts
Durable work state
A request becomes a task with attempts, messages, artifacts, dependencies, results, and provenance stored by the local control plane.
- Chronological task conversation and activity
- Follow-up context stays with its task
- Questions can wait for a human
- Configurable daily and monthly budget policies
Background automation
Queue work now, run it on a cron schedule, arm a one-shot wakeup, or move it through a reusable multi-harness procedure.
- Task queue survives control-plane restarts
- Recurring schedules and one-off follow-ups
- Sequential multi-agent workflows
- Durable GitHub review waits without keeping a worker alive
BRING YOUR HARNESS
Keep the intelligence. Add somewhere durable to work.
XpressClaw is an ACP client and control plane. It does not replace the selected product’s reasoning, tools, permissions, or subagents.
14 built-in ACP harnesses in the current release catalog Detected tools appear first during setup; each runner image stays product-specific.
CodexClaude CodeOpenCodeGitHub CopilotJunieKimi CLIpi ACPQwen CodeClineCursorGLM AgentGrok BuildKilo CodeMistral Vibe + Custom ACP runner
Harness-owned intelligence
Use the native product, not an imitation of it.
ACP, not another wrapper loop
The harness owns planning, tool use, permissions, and subagents. XpressClaw owns durable work and presentation.
Advertised controls in the UI
After discovery, supported commands, modes, models, reasoning levels, and MCP choices appear where work is composed.
Native configuration can come with you
Trusted runners can reuse standard logins, settings, skills, plugins, hooks, and custom agents from the host.
One Agent, replaceable harness
Keep project context stable while execution products evolve.
Durable Agent context
An Agent ties together a name, workspace, task history, runtime configuration, and the currently selected harness.
Custom ACP runners
Provide a container image and command for another coding harness that speaks ACP over standard input and output.
Multiple products in one procedure
Bind workflow roles to different Agent contexts so implementation, review, and follow-up can use different harnesses.
Explicit runtime boundary
Power is visible and opt-in.
Container isolation is required
Worker environments run through Docker or Podman and receive only configured workspaces, volumes, and environment values.
Host-engine access is a trusted mode
Optional Docker, Compose, and Buildx access gives the runner control of the host engine and is clearly treated as privileged.
Provider boundaries stay honest
Control-plane data stays local; prompts and code sent to a cloud harness follow that provider’s terms and configuration.
THE BROADER ARCHITECTURE
Persistent agents are bigger than coding.
Native harnesses are excellent specialized execution engines. Persistent coworkers add long-lived context, memory, judgment, objectives, and delegation. The long-term XpressClaw architecture is designed for both to coexist.
An agent is a graph, not a prompt.
SHIPPED TODAY
Native agent harnesses
Codex, Claude Code, OpenCode, and other ACP products bring focused execution loops that are already excellent at software work.
- Fast specialist execution
- Native tools and subagents
- Replaceable per Agent context
- Ideal for focused jobs
ARCHITECTURAL DIRECTION
Persistent Xpress / Xaibo agents
Longer-lived coworkers can combine reasoning, memory, review, perception, tools, procedures, knowledge, objectives, and identity through explicit modules.
- Organization and repository context
- Long-lived memory and learning
- Objectives and social identity
- Delegation to specialist harnesses
Xaibo
01Primary reasoning
02System 2 / deep reasoning
03Memory & consolidation
04Pre-action review / safety
05Outcome review / learning
06Tools, skills & procedures
07Perception & execution
08Objectives & identity
What ships in XpressClaw 0.2 beta
The public beta currently exposes native ACP harnesses, durable Agent contexts, tasks, schedules, workflows, task history, and retained project runtimes. Xaibo coworkers and shared conversation rooms are not yet available in the XpressClaw product UI.
COORDINATION MODEL
Chat coordinates. Tasks endure.
XpressClaw is not defined by a Kanban board or a fixed workflow builder. Tasks are the durable execution substrate; the public beta’s current boundary is explicit below.
SHIPPED TODAY
Start work from the Agent UI
In the public beta, New Work or an Agent workspace accepts an outcome and creates a task for the selected Agent and harness.
SHIPPED TODAY
The task carries execution state
Attempts, tool activity, questions, follow-ups, artifacts, and results persist. Schedules and workflows put work onto the same substrate.
PRODUCT DIRECTION
Shared conversations become the coordination layer
The intended flow is conversation → accepted responsibility → task → background execution or delegation → meaningful progress back to the conversation. Shared rooms and channels are not in the current public beta.
BUILT IN PUBLIC
Open source should feel like a project, not a teaser.
XpressClaw is useful as a local control plane on its own. The public repository contains the runtime, UI, runner definitions, workflow examples, and docs. Inspect the boundary, extend a runner, join Discord, or send a pull request.
0.2Beta release line
14Built-in harnesses
3Desktop platforms
GPL-3.0License
A PRECISE TRUST BOUNDARY
Local control does not mean every selected model is local.
Configuration, task state, and timelines are stored by the local control plane. A cloud harness or model still receives the context you send under its provider terms. Host credentials are mounted only when enabled and only into runner images you choose to trust.
XPRESSCLAW + XPRESS AI
One direction, two ways to operate it.
XpressClaw is the open-source local foundation and experimentation surface—not a deliberately crippled demo. Xpress AI is the managed platform for organizations that want persistent coworkers without operating the surrounding infrastructure.
OPEN SOURCE · LOCAL
XpressClaw
For developers who want to own the control plane, bring native harnesses, inspect the runtime, and build durable agent work on their own machine.
- Local control-plane state
- Native ACP harnesses
- Tasks, schedules, and workflows
- Retained Docker or Podman runtimes
MANAGED PLATFORM
Xpress AI
For organizations that want hosted persistent AI coworkers with managed collaboration, infrastructure, integrations, and governance.
- Hosted persistent agents
- Shared conversations and channels
- Managed knowledge and integrations
- Cloud desktops, communications, and governance
START LOCAL. KEEP CONTROL.
Give one agent one durable job tonight.
Download the beta, choose a project folder and supported harness, then send the first outcome from New Work.