Home · Mewbo

Documentation

Introducing Mewbo¶

Mewbo is an open stack for handing agents to your engineers and operators, inside the permissions you set. It plugs into the tools and systems you already run. Automation, wiki, search and apps stand on one harness, whichever model you bring.

Quickstart → Install via Docker API reference

Parallel sub-agents Claude Code and Codex compatible Self hosted, any model Permissions, plan mode, sandboxed shell Answers cited to a path and line

The Mewbo Console home listing recent sessions — Your sessions at a glance

A Mewbo task in the console, broken into steps with tool calls and results — Inside a task, step by step

Interactive widgets rendered inline in a Mewbo conversation — Interactive widgets, inline in chat

An Agentic Wiki overview page with a runtime flow diagram and an Ask MewboWiki box — Agentic Wiki: documentation grounded in your code

An Agentic Wiki answer to a question about a project, with a summary card, an expandable Cited Sources list, a Retrieval details panel naming the pages accessed and the model used, and the answer itself carrying inline citation chips — Ask the wiki, and every claim carries its source

The Agentic Search landing page with workspaces scoped to connected sources — Agentic Search: workspaces over your connected tools

Agentic Search results: one ranked list across connected sources with a synthesised overview — One ranked list across every tool, topped by a synthesis

The Mewbo plugins page with installed plugins and marketplace listings — Plugins and a marketplace to extend any session

A live Mewbo App called LLM Model Compare, with a filter rail, a bar chart ranked by coding score, and a right rail showing health, recent runs, daily pipelines and versions — Agentic Apps: a sub-agent writes it, then it runs on a schedule

Two Mewbo settings panes side by side, one listing installed plugins and what each contributes, the other listing reverse invocation triggers with their cron expressions, fire counts and next fire time — Reverse invocation: a trigger starts a session with nobody watching

Choose your surface¶

A coding agent serves one person. Mewbo deploys once and serves every role, on the surface each person already works in.

Terminal Hand off the refactor, the migration or the flaky test you have been putting off, and stay in the shell while it works. It runs on your machine, against the branch you are already on. Web console The whole team works here. Each role sees only its granted tools, escalation is one request away, and a supervised run becomes a repeating trigger. Android Make Mewbo the assistant your phone answers to. Ask out loud from any screen or answer a permission prompt on the move, against the same deployment your team already runs. Team chat Mention Mewbo in a channel and it works in the open. The thread is the record, so the next person picks it up without asking anyone what happened. Email Send it work the way you would send a colleague work. Reply in the thread to push it further, from a phone, a laptop or an inbox rule. Home Assistant Ask which rooms are free, why a machine stopped, or what tripped overnight. Every entity your site exposes is now something you can ask.

How it works¶

Every product on this page runs the same three steps underneath.

You describe an outcome¶

Ask in plain English on whichever surface is closest. Calls run as the model generates them, and anything the permission policy does not match comes back to you as a prompt. Switch to plan mode for unfamiliar or destructive work and the run stops at a drafted plan instead. Reject it with feedback and a revised plan comes back.

Mewbo delegates in parallel¶

The root agent spawns a sub-agent for every piece that can run at once. A test run, a search, a refactor and an MCP call all execute in parallel. A stalled child is caught by the hypervisor, which injects a correction between tool steps. A budget that runs out buys a turn to wrap up rather than a bare halt.

You get a synthesised answer, not a pile of logs¶

Each sub-agent returns a structured result with its status, summary, warnings, files touched and checks against the acceptance criteria. The root synthesises those into one answer, backed by a transcript of every tool call, permission prompt and compaction. Fork any message to replay that branch on a different model.

Go deeper

Architecture Overview covers the tool use loop, the hypervisor and the concurrency lifecycle.

Highlights¶

Agentic Wiki Every repository gets documentation that stays current, and a place to ask questions about it. A new teammate asks how something works instead of who wrote it, and the answer comes grounded in the code and cited so it can be checked. Agentic Search One search across every system your team has connected. Ask where a decision landed and sub-agents query your trackers, chats and repositories at once, merging what they find into a single ranked answer that names its sources. Control over what actually runs Giving a team an agent is a permissions decision. Your identity provider says who people are, roles and teams say what each may run, a policy checks every tool call, and the shell sits in a kernel sandbox. All four are set in one place. Agentic Apps Describe an app in a sentence and an agent builds it over your live systems, from the data model to the interface to the pipelines that keep it fresh. It runs in your deployment under your permission rules and repairs itself when a feed breaks. Mewbo as an MCP server Claude Code, Codex or Cursor calls Mewbo as a tool. Send a whole task to your deployment without leaving the editor, or ask the wiki about a repo you have never opened, while your session keeps its context on the task at hand. Structured Outputs Give your own software an agent without building one. Send a question and the JSON Schema you need back, and a managed session researches your code and tools, then returns an answer that fits it exactly.

Already using Claude Code or Codex?¶

Point Mewbo at a project and it loads the MCP servers, skills, plugins and instruction files you already have.

MCP servers Both the Mewbo servers and the Claude Code / VS Code mcpServers schemas are accepted at project and user scope.

Skills SKILL.md files in ~/.claude/skills/ or .claude/skills/ activate exactly as authored, with the Agent Skills standard.

Plugins & marketplaces Claude Code plugin manifests install from any marketplace that already serves them, with commands, agents and tools carried over.

Project instructions CLAUDE.md, AGENTS.md, and .claude/rules/*.md all load hierarchically on session start.

Session tools A session_tools array in plugin.json contributes stateful tools that live for one agent.

What you can do¶

Workspace & execution

Built-in tools: read, edit, shell, list
Widgets: interactive UI inline in chat
External MCP tools: servers Mewbo calls
Web IDE: per-session code-server
Code intelligence (LSP)

Knowledge & discovery

Agentic Wiki: source-grounded repo docs
Knowledge graph of the codebase
Agentic Search across connected MCPs
Agentic Apps: a data layer and interface over any of it

Composition & delegation

Control & safety

Identity, roles and teams
Plan mode: review before execution
Permissions and hooks
Sandboxed execution: a kernel confined shell
Policies: semantic gate-checks on tool calls
Monitors: session-wide behavioural guardrails

Under the hood

Architecture overview
Session runtime
Token usage and budgets
Mewbo as an MCP server: agents that call in
Full API reference

From install to production¶

Install

Configure

Use

Deploy

Extend

Keep learning¶

GitHub repo Source, issues, and releases. Architecture deep-dive The tool use loop, the hypervisor, and the lifecycle. Troubleshooting Common errors and how to diagnose them. Changelog Release notes and upgrade guides.