MindFlock
Turns tickets assigned to you into a private flock of AI coding agent sessions, each in its own git worktree on your own machine.
- Source repo
- MindFlock/MindFlock
- Stars
- ★ 15
- Last updated
- today
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 59/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeOpenAI API · Claude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A solo developer whose Jira or Linear queue is the backlog, who wants assigned tickets to start sessions with nothing typed.
- Not a fit if
- Teams that need work to continue while the laptop is shut
- Anyone unwilling to send code to any vendor model (short of a local model)
- Mac-only users who insist on a native client rather than Electron
- Website
- mindflock.ai ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v0.14.23 (2026-11-20)
- Source review
- 59/100 · Major gaps
What does this agent do, and when should you use it?
MindFlock is a local-first workspace for running many coding agents side by side. It polls GitHub Issues, Jira, Linear, Shortcut and Asana for work assigned to you, plus your own pull requests that came back with review comments, and for each one creates an isolated git worktree on a feature/<ticket>/<slug> branch, installs dependencies and launches an agent CLI already seeded with the ticket title, description, mined acceptance criteria and comments. Internals are split into backend.session (session lifecycle, worktree management, tmux/ConPTY), backend.web (FastAPI server plus the UI the Electron desktop app renders), backend.ticket_ingestion (polling, filtering, dedup) and backend.providers (TOML-declared agent CLIs with hook-based activity detection, local-model routing and cost tracking). Nothing is cloned to a vendor sandbox: the engine binds 127.0.0.1 by default, ships no telemetry and keeps tracker tokens in ~/.mindflock/settings.json at mode 0600. By default nothing is committed, pushed, opened or merged without your click.
The ingestion pipeline polls each configured tracker and GitHub roughly every 20 seconds, deduplicates against persisted state, and provisions a session per ticket or per PR: backend.session creates the git worktree and branch, runs stack-specific setup (uv sync --all-groups and pre-commit install for Python/uv repos, otherwise declared setup_commands), then launches the agent CLI configured for that source with the ticket's title, description, acceptance criteria and comments as the prompt. Sessions run in detached tmux (or a ConPTY session host on native Windows); events stream over WS /api/events and provider-defined hooks classify each agent as working, idle or needs-input. In the desktop app or the /m phone UI you read the Diff tab and drive the work home with one click each: commit (executed in the session's own terminal so hooks are visible), push (plain git push over your existing remote), Open PR and Merge, falling back from gh to a GitHub token to a prefilled compare URL. Extra surfaces include New → Ticket (describe a task, file a real ticket), duplicate merging from Intake, per-source opt-in tracker writes, Verify checklists, notifications via ntfy, team runs with budgets, and an extension/hook system.
- A solo developer whose Jira or Linear queue is the backlog, who wants assigned tickets to start sessions with nothing typed.
- A team that wants several agents editing one repository in parallel without the code being cloned to a vendor VM.
- A maintainer whose PRs come back with inline review comments and wants each round of unresolved comments to become a new session on the PR's branch.
- A developer without a cloud sandbox budget who wants to run the whole ticket-to-PR loop against Ollama or LM Studio on a local model.
- Someone moving between an office desktop and a laptop who wants sessions from all machines in one sidebar and to approve commits from a phone.
- A reviewer who needs to fence off directories with Keep out / Only here zones so agents cannot edit outside agreed areas.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -LsSf https://raw.githubusercontent.com/MindFlock/MindFlock/main/install.sh | shirm https://raw.githubusercontent.com/MindFlock/MindFlock/main/install.ps1 | iexuv tool install "mindflock[web] @ git+https://github.com/MindFlock/MindFlock@v<version>"
- Environment variables
MINDFLOCK_NTFY_TOKEN- Config files
config.toml- Runtime
- Python 3.12+
- Install notes cover
- Windows · macOS · Linux · WSL
How do you install or deploy this agent?
The default path is a release installer: MindFlock-Setup.exe on Windows, MindFlock.dmg on macOS (self-signed, so the first open needs System Settings → Privacy & Security → Open Anyway, or the de-quarantine command below), MindFlock.AppImage on Linux. First launch installs the engine and required tools itself.
xattr -dr com.apple.quarantine /Applications/MindFlock.appFor a terminal-only install, run the install script on Linux, macOS or WSL; on native Windows use the PowerShell twin:
curl -LsSf https://raw.githubusercontent.com/MindFlock/MindFlock/main/install.sh | sh
mindflock init
mindflock serveirm https://raw.githubusercontent.com/MindFlock/MindFlock/main/install.ps1 | iexYou can also pin a release tag with uv or pipx, or build from a clone:
uv tool install "mindflock[web] @ git+https://github.com/MindFlock/MindFlock@v<version>"
git clone https://github.com/MindFlock/MindFlock
cd MindFlock
uv sync --group webRequirements: Python 3.12+, tmux ≥ 2.4 (not needed on native Windows, which uses a ConPTY session host instead), git ≥ 2.17, Git for Windows on native Windows (its bash.exe runs each session's launch script), a coding-agent CLI (claude by default), and a git remote you can already push to over SSH or HTTPS.
How do you use this agent?
Start the server and open the UI, or create a session straight from the terminal:
cd ~/code/your-repo
mindflock init
mindflock serve # binds 127.0.0.1 by default; open http://127.0.0.1:8765
mindflock new . -p "fix the failing tests"
mindflock ls
mindflock attach webapp
mindflock msg webapp "also update the changelog"To let a tracker start sessions, open Intake → Tickets (Alt+I) in the desktop app, click + Add ticketing source, enter a token and target repo, then flip the tab's Automated ingestion switch. GitHub Issues needs no configuration — it reuses your existing gh auth login. For phone access, run tailnet mode and scan the QR code printed at startup:
mindflock serve tailscaleHeadless setups can use config.toml or run the pipeline standalone:
python -m backend.ticket_ingestionPhone push requires the ntfy app and a generated topic in Settings, or environment variables:
export MINDFLOCK_NTFY_TOPIC=mindflock-xTPq…
export MINDFLOCK_NTFY_SERVER=https://ntfy.example
export MINDFLOCK_NTFY_TOKEN=tk_…What are this agent's strengths and limitations?
- Your repository never leaves your disk: no MindFlock cloud, no account, no vendor sandbox, engine bound to 127.0.0.1 by default, with a full outbound-network inventory documented.
- Tracker ingestion is the differentiator: five ticket sources plus your own reviewed PRs each become an isolated worktree with a context-seeded agent, and a ticket is never worked twice.
- You keep merge authority: nothing is committed, pushed, opened or merged until you click, commits run in the session terminal so hooks are visible, and depth only increases if you set Goes as far.
- Provider-agnostic by configuration: every agent CLI is a TOML file (Claude Code, Codex, Antigravity, aider, OpenCode, Cline, Goose bundled), each tracker source can pick its own agent, and hook-based working/idle/needs-input detection plus token and cost tracking apply to all of them.
- Wide surface for supervision: Electron desktop app, /m phone UI with Diff and voice, a CLI driving the same sessions, tailscale QR access, ntfy push, and an MCP server so agents can message each other.
- Setup is not a single clean command: tmux, git and an agent CLI must be present, tracker tokens must be pasted per source, and the macOS build is self-signed so Gatekeeper blocks the first launch.
- The agent still sends code to its own vendor (Anthropic for claude, OpenAI for codex); closing that requires a local model, and Claude Code has no local route at all.
- Ingestion polls instead of listening for webhooks, so a new ticket can take up to ~20 seconds to appear, and review sessions deliberately ignore PRs younger than 15 minutes.
- Dependency installation is only auto-detected for Python/uv repos; every other stack must declare setup_commands, which is an adoption cost before the first session runs correctly.
- The headline 5.9× / 96%-tests figures come from one developer's own repository and are stated as such; they are not a validated team-level benchmark.
How does this agent compare with similar options?
The README contrasts two families of tools. Cloud ticket-to-PR products — Devin, Google Jules, GitHub Copilot coding agent, Cursor background agents, Atlassian Rovo Dev — also start from an assigned ticket but clone the repository into a vendor-operated sandbox or VM; MindFlock's bet is that the code never moves. Among local orchestrators, Claude Squad is a mature multi-agent terminal TUI, Conductor is a native macOS app, and Claude Code's built-in Agent Teams + worktrees is free and already installed — all of them require you to start each session yourself, whereas MindFlock lets the ticket queue do it. The README also lists Cyrus, no_human, Frink and Open Session as open-source projects that start work from tickets on hardware you control but rely on webhooks, OAuth apps or public ingress; MindFlock polls the queue already assigned to you instead.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| MindFlock This agent | 59 · Major gaps | Desktop appFree + model costs | ★ 15 | today | Python | Codex · Claude Code |
| Pi Agent Desktop | 78 · Good | — | ★ 353 | 9d ago | TypeScript | OpenAI API |
| Poracode | 67 · Some gaps | Desktop appFree + model costs | ★ 123 | today | TypeScript | ChatGPT · Claude Code |
| Verde | 46 · Major gaps | Desktop appFree + model costs | ★ 36 | 1d ago | Zig | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
README and SECURITY.md give unusually detailed data-flow, token storage (0600), Origin/Host guards, and default no-auto-push/merge behavior, so transparency is strong. However, agents default to --dangerously-skip-permissions while ingesting untrusted ticket text, weakening least privilege; rollback relies only on git worktrees and manual action with no explicit undo; source attribution is self-asserted and the publisher is unverified.
Code shows singleton locking, exit records, and decent failure messaging, but dependency pins are loose (aiohttp, fastapi without upper bounds) and no executed verification is available, so dependency availability is only partially supported.
README covers multi-platform, multi-tracker, multi-CLI and local-model scenarios and states capability boundaries (claude has no local route, no auto-merge), but trigger precision and real environment fit rest on documentation claims, not demonstrated runs.
Documentation is well-structured with complete install steps, honest known limitations, and a full Apache-2.0 LICENSE; but versioning appears only in pyproject with no CHANGELOG, and maintenance responsibility/publisher identity is unclear.
Output usability (diff, PR, cost tracking) and marginal value (local execution, parallel agents) are described, but cost-benefit rests on the author's own self-reported data and admits heavy token consumption and increased engaged time, so evidence is insufficient for a high score.
README provides methodology and reproducible definitions and even corrects prior inflated figures, separating fact from inference well; but core benefit numbers come from the author's own repository without independent corroboration.
- Agents default to --dangerously-skip-permissions while ingesting untrusted ticket text, creating prompt-injection risk; evaluate and consider disabling it.
- Dependencies lack upper-bound pins, so supply-chain risk should be locked down independently.
- Benefit figures come from the author's own repository without independent verification and should not be taken at face value.
- Publisher identity is unverified and the maintenance/update path is unclear.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
Added - Disk cleanup: a garbage collector for leftover worktrees. Agents leave worktrees behind fast (one machine had 141 folders, 118 GB, most from sessions closed weeks ago). MindFlock now keeps an inventory of every worktree and session clone it made — plus the other worktrees of the same repositories, which it lists but never removes — and classifies each one: in use (a session, sleeping or paused included, one closed recently enough to reopen, a route, run or queue), uncommitted changes, commits that are in …View on GitHub ↗ -
Release notes excerpt
Added - Updates take care of themselves. One setting, Settings → Advanced → Updates, the same in the browser and the desktop app: Automatic (a new version installs by itself and the server restarts onto it when nobody has typed for a minute, or after 30 minutes; sessions keep running), Ask me (one offer per version, with an Update button right on the toast and the bell row) or Off. Ask me is the default; a desktop app that had automatic updates on keeps them. Also mindflock update --auto on|ask|off and mindflock update --check. …View on GitHub ↗ -
Release notes excerpt
Added - To-do is its own extension. A To-do bar in the sidebar (show or hide it in Customize → Sidebar) with a count of what is open, Open and Add. The list opens as a window in the grid like any session: what needs you first, grouped by session, each with a chip that takes you to that session; drag to reorder; High / Normal / Low; details per item (links, and code blocks with a Copy button); a due time; and a History tab of what you finished or removed, with Restore. An item you check off stays …View on GitHub ↗
FAQ
Does my code get uploaded to a MindFlock server?
Can I use it without a subscription or API credits?
Will it commit, push or merge on its own?
How is remote and phone access protected?
What is left behind if I uninstall?
Is MindFlock free?
Does MindFlock work on Windows?
irm https://raw.githubusercontent.com/MindFlock/MindFlock/main/install.ps1 | iex.Is MindFlock open source?
Is MindFlock safe to use?
Is MindFlock still maintained?
What is MindFlock's official website?
Thanks — we use this to improve the page.