Flock Multi-Agent Workbench
Build, run, and supervise multi-agent automation through a visual desktop workspace.
The README says that consequential file writes, Bash commands, and configuration changes require approval, and it describes container isolation, VNC inspection, SQLite history, and encryption utilities. Cargo includes ring for AES-256-GCM and explicitly pins rustls-webpki for identified RUSTSEC advisories, supporting partial confirmation, secret-handling, and dependency-security credit. Deductions apply because the agent also has full filesystem, shell, browser, MCP, and unattended scheduling capabilities without documented permission scopes, network or API-key flows, key lifecycle, operation-by-operation approval rules, or recovery procedures. Workflow history is not demonstrated rollback for external effects. Langgraph-rust is attributed, but the reviewed repository is Onelevenvy/flock while Cargo points to iOfficeAI/flock, and publisher identity is unknown, leaving attribution inconsistent.
The Rust, Node, Tauri, React, and LangGraph description is partly consistent with the Cargo workspace and release workflow, while dependencies and the cross-platform build matrix are concrete. Deductions apply because the README permits Rust 1.77.2 while the workspace declares edition 2024, which may imply an inconsistent toolchain requirement; the README uses npm while releases use pnpm with a non-frozen lockfile. No supplied file demonstrates user-facing error, denial, timeout, recovery, or diagnostic messages, so failure_messages receives zero.
The README identifies desktop multi-agent orchestration, visual workflows, browser automation, human intervention, and scheduled automation as target scenarios. Multiple model providers, MCP, hot-reloaded skills, and three desktop platforms provide reasonable ordinary-use adaptability. Deductions apply because unsupported scenarios, resource needs, container and VNC prerequisites, and platform differences are not documented. Approval triggers are described only as 'crucial actions,' while scheduled jobs may operate unattended, so pause, inheritance, and interaction-bypass behavior is not precisely bounded.
Quick navigation, feature sections, an architecture diagram, and a crate responsibility table provide strong information architecture; installation and test commands offer a basic development path. Apache-2.0 is consistently represented by metadata, Cargo, README, and a complete LICENSE, justifying full license credit. Version 0.2.0, tag-driven manifest synchronization, and Release Drafter provide a credible versioning mechanism. Deductions apply for the absence of an FAQ, end-to-end examples, troubleshooting, and known limitations; npm versus pnpm and conflicting repository-owner paths weaken naming and setup stability. A personal implementation statement and repository address do not clearly establish maintainers, support channels, or security-update responsibility.
If implemented as described, the visual editor, execution history, streaming, human nodes, and VNC takeover could make agent results inspectable and provide meaningful orchestration value beyond a basic chat client. Deductions apply because the evidence consists mainly of feature claims and referenced screenshots, without concrete output examples, quality measures, or implementation excerpts establishing that results are directly usable. Claims of major performance, reliability, and experience improvements are unsupported. The project is described as free and zero-configuration, but API keys and possible model, container, browser, and continuous scheduling costs are not analyzed.
Some architecture, crate, version, dependency, release-platform, and license claims are corroborated across Cargo and workflow files, providing limited cross-source verification. Deductions apply because most security, feature, node-count, approval, and performance claims appear only in the README without corresponding code, configuration, or test evidence. Promotional conclusions are not clearly separated from observable facts. Conflicting repository URLs, package managers, and minimum toolchain information further reduce traceability.
- Before granting real filesystem, shell, browser, or MCP access, verify that approval gates cover every high-impact operation and that unattended schedules cannot bypass them.
- Confirm the encryption-key source, at-rest API-key location, log redaction, export, and deletion policies; the supplied evidence establishes only encryption-related claims and dependencies.
- Resolve the actual minimum Rust version and whether npm or pnpm is authoritative, then use a locked dependency installation.
- Clarify the relationship between the Onelevenvy and iOfficeAI repository paths, the current maintainer, and the security-update channel.
- Container isolation, VNC, Playwright behavior, and cross-platform support were not executed or independently validated in this static assessment.
What does this agent do, and when should you use it?
Flock is a desktop multi-agent harness built with Rust, Tauri, and React, with graph execution powered by langgraph-rust. Its ReactFlow editor connects agent, model, routing, code, human-intervention, plugin, and data-extraction nodes while exposing streaming runs and execution history. The Rust backend is divided into flock-agent, flock-workflow, flock-tools, flock-skills, and flock-core, with SQLite checkpointing and Tauri IPC linking it to the interface. Agents can work with files, run Bash, call MCP tools, control a Playwright browser, and execute inside isolated environments whose desktops are visible over VNC. It runs as a local desktop application and supports OpenAI, Gemini, Anthropic Claude, AWS Bedrock, and Ollama/local models, making it suitable for adopters who want control over both execution and model choice.
A user creates a graph on the ReactFlow canvas, beginning with start and combining llm, agent, classifier, ifelse, code, human, plugin, parameter_extractor, and answer nodes for inference, routing, transformation, intervention, tool exposure, extraction, and final delivery. flock-workflow compiles workflow JSON into a LangGraph AST; flock-agent executes it through the langgraph-rust state machine and persists state through a SQLite Checkpointer. During a run, flock-tools can read and write files, execute Bash, perform grep/glob operations, invoke connected MCP Server tools, or use Playwright, xdotool, X11, and VNC inside an isolated environment. The interface requests approval for consequential actions such as writing files, running scripts, and changing configuration, while human nodes provide explicit workflow breakpoints. Results are delivered through the answer node, and the UI retains workflow versions, execution paths, and historical states; standard cron syntax can also schedule recurring runs.
- An automation team can turn model calls, conditional routing, code transformations, and approval gates into a visual pipeline that can be inspected and debugged.
- A developer or operator can let an agent organize local files or run maintenance commands while retaining approval over sensitive file, Bash, and configuration actions.
- A team producing recurring reports, aggregating files, or performing system maintenance can schedule unattended runs through the native Cron scheduler.
- A research or operations user can automate scraping, form interaction, and virtual-desktop tasks with Playwright and xdotool, then take over through VNC when a captcha or other obstacle appears.
- A group with several specialized agents can use Flock to detect and orchestrate them as a team network while monitoring each execution step.
- An organization with existing MCP servers can connect them once and expose their schemas and tools across assistants and workflows.
What are this agent's strengths and limitations?
- Combines multi-agent execution with a ReactFlow visual builder offering ten documented node types, streaming execution, workflow versions, and historical-path debugging.
- Interactive approvals, human-node breakpoints, and VNC takeover provide distinct control points for sensitive or long-running automation.
- Its isolated environment combines Playwright, xdotool, X11, and VNC, covering both browser automation and observable graphical-desktop control.
- Documented support for OpenAI, Gemini, Anthropic Claude, AWS Bedrock, and Ollama/local models avoids dependence on a single model provider.
- Shared MCP integrations and standard Cron scheduling cover both interactive tool use and recurring unattended workflows.
- Source-based setup requires Rust 1.77.2+, Node.js 18.x+, npm, Tauri frontend tooling, and several Rust crates, creating more operational overhead than a conventional chat client.
- Cloud-model routes require users to supply API keys and remain subject to the chosen provider's network availability and billing; only the local-model path avoids that dependency.
- Arbitrary JS/Python, Bash, filesystem writes, browser control, and desktop input are powerful permissions; adopters still need an appropriate approval and container-security policy.
- The supplied documentation does not identify prebuilt installers, exact supported desktop OS versions, production packaging steps, or a complete UI configuration walkthrough.
- The application was fully rewritten from a Python, LangGraph, LangChain, and FastAPI backend to Rust and Tauri; users depending on legacy behavior or extensions must assess migration work, with the old implementation retained only in legacy/python.
How do you install or deploy this agent?
Install Rust 1.77.2 or later, Node.js 18.x or later, and npm. Then run:
git clone https://github.com/Onelevenvy/flock.git
cd flock/flock-ui
npm install
npm run tauri devTo test the Rust backend, run cargo test --workspace from the repository root. First use also requires an API key for OpenAI, Gemini, Anthropic Claude, or AWS Bedrock, unless you select Ollama or another local model. The supplied material does not document a prebuilt-package command, OS-specific prerequisites, or a production packaging and release procedure.
How do you use this agent?
Start the desktop development application with npm run tauri dev, then paste an API key for a supported provider or configure Ollama/a local model. Use the built-in agent directly, or create a ReactFlow workflow beginning with start, connect llm, agent, routing, code, human, plugin, or other supported nodes, and define final delivery with answer. Watch the streaming execution state and inspect any approval prompt before allowing file writes, Bash commands, or configuration changes. Browser and desktop automation can run in an isolated environment, with VNC available for mouse and keyboard takeover. Connect an MCP Server when external tools are needed, and use standard cron syntax for recurring tasks. The supplied material does not provide exact UI click paths or a complete sample workflow.
How does this agent compare with similar options?
Compared with the traditional chat clients described in the README, Flock adds a ReactFlow workflow builder, filesystem and shell operations, autonomous multi-step LangGraph execution, Cron scheduling, and multi-agent collaboration. That broader scope also requires a local development runtime, model credentials, and deliberate management of execution permissions and sandbox boundaries.