AgentForeman

Approve, steer, and stop every Claude Code agent from one page.

AgentForeman is a local web page for the Claude Code sessions on your Mac. It shows sessions in VS Code panels and in terminals, their subagents, and your recent Codex threads. It shows which sessions wait for you, and what every agent does now. With its hooks on, you can answer and steer those agents from the same page. It runs on your Mac, and only you can open it.

Why AgentForeman

AgentForeman works on the sessions that you already run, and it acts on all of them from one page:

• Works on the sessions you already run. It finds every Claude Code session on your Mac. You start nothing through it.
• Approve with the whole command and its diff. The card shows the full command, or the diff of an edit.
• Approve every matching call at once. One click answers every waiting call of the same kind, across sessions.
• Allow rules from your approvals. It suggests a rule after repeat approvals. Rules start off, and they fail closed.
• Steer a running session. The agent reads your note at its next tool call, and keeps working.
• Stop one subagent or a whole session. The stop takes effect at the next tool call.
• Warn when two sessions edit one file. The second session gets a warning, or Claude Code asks you first.
• Show split-pane teammates. Agent Teams teammates in their own tmux pane show under their lead.
• Local, with nothing to build. Only the Python standard library. No account and no cloud service.

Compared with other tools shows how 59 other tools differ.

Install

With pipx:

pipx install agentforeman

With uv, without an install:

uvx agentforeman --open

From source:

git clone https://github.com/xmpuspus/agentforeman
cd agentforeman
python3 -m agentforeman --open

Quick start

1. See it on sample sessions first. The demo changes no file outside a temporary folder:

   agentforeman demo
   

   Open the link that it prints. Press Ctrl+C to stop the demo.

2. Watch your own sessions:

   agentforeman --open
   

   The page opens in your browser, and the browser stays signed in. To open it again later, or in another browser, run agentforeman open. Until you turn on the controls, the page only reads your sessions.

3. Turn on the controls, so you can approve, reply, steer, and stop from the page:

   agentforeman install
   

   It saves a backup of ~/.claude/settings.json, then adds four hooks. Running sessions get them without a restart. To remove them, run agentforeman uninstall.

serve is the default command, so agentforeman --port 4330 works too. Its options:

• --port: the port of the page. The default is 4320.
• --open: open the page in your browser.
• --claude-dir: where Claude Code keeps its session files. The default is ~/.claude.
• --codex-dir: where Codex keeps its thread files. The default is ~/.codex.
• --control-dir: the folder that the page and the hooks share. The default is ~/.agentforeman.
• --dry-open: the button Open in VS Code returns its commands and runs nothing.

open takes --port, --control-dir, and --print. With --print, it prints the sign-in link instead of opening it.

install and uninstall take --settings and --control-dir.

Features

See which sessions wait for you

The Overview counts the sessions that wait for you, the running and idle sessions, the recent tool calls, and the live sessions. The table Needs attention lists each session that waits for you, oldest first. The badge Needs input means that the session ended its turn with a question. The badge Awaiting approval means that a tool call waits for your permission. The browser tab title starts with the count, for example "(3) AgentForeman".

Approve or deny a tool call

The card shows the whole command that waits, with its line breaks. For an Edit or a Write, it shows the diff. Click Approve or Deny. You can also answer in VS Code, the terminal, or the Claude app. The first answer wins, and the card goes away.

Approve similar calls at once

When other sessions wait for the same kind of call, the card shows Approve all N. Click it, then click Confirm within 4 seconds.

Always allow rules

The card names a rule for the call, for example Always allow Bash(make eval). Click it to add the rule. Adding a rule never turns rules on. Turn rules on with the switch on the Rules page.

While rules are on, a matching call runs at once, also while the page does not run. The Rules page lists each rule with its hits, and the last 20 calls that rules allowed. A rule never matches a command with $, a redirect, or a part that no rule covers. Such a call comes to you.

Reply to a session

When a session ends its turn, its panel shows a reply box. Type your reply, then click Send reply or press Cmd+Enter. The session wakes up and reads your text.

Steer a running agent

A running session, subagent, or teammate has a note box in its panel. The agent gets the note at its next tool call, as extra context, and keeps working. The panel shows when the note arrives. A chip such as "Repeating: Read config.toml x5" marks an agent that makes the same call again and again.

Stop a subagent or a session

Each running subagent has a Stop button. A running session has the button Stop session. Click, then confirm within 4 seconds. The agent stops at its next tool call. A command that already runs keeps running until it ends.

Same-file guard

When two sessions edit one file, the Overview shows the file and both sessions. The guard warns the second session before its edit. On the Rules page, choose Off, Warn, or Block. The mode Block makes Claude Code ask you before such an edit.

Subagents and teammates

The Subagents page groups subagents by session, for the last 6 hours. A subagent that another subagent started sits under its parent. Agent Teams teammates show here too, also the ones in their own split pane.

Session panel

Click a session row to open its panel. The panel shows a timeline of tool calls, a trace on one clock, the subagents, and the details. The button Open in VS Code brings the window of that session forward. The button Open in Claude app opens its Remote Control page on claude.ai.

Search and keys

Search finds sessions by title, folder, message, file, command, or subagent. The project menu shows one project folder. Keys: / searches, j and k move, Enter opens, o opens VS Code, and Esc closes.

Activity and Codex

The page Activity shows tool calls per minute for the last 30 minutes, with an event log. The page Codex lists Codex threads from the last 48 hours. The theme follows your system until you pick one. With Alerts on, a desktop alert shows when a session starts to wait for you.

Compared with other tools

We read the README or the docs of 59 tools on 4 October 2026. Most of them fall into three groups:

• Dashboards show sessions, costs, and logs. A few of them also approve tool calls.
• Session managers start agents for you, often each in its own git worktree. Your sessions run inside them.
• Remote approvers send permission prompts to your phone or a chat app.

AgentForeman works on the sessions that you already run, and it acts on all of them from one page. The table shows the tools that have at least one of these features. A dot means that the README or the docs do not mention the feature. It does not prove that the tool lacks it.

Tool	Finds the sessions you run	Approve a call	Note to a running agent	Stop one subagent	Same-file warning	Rules from approvals
AgentForeman	Yes	Yes, with the diff	Yes	Yes	Yes	Yes
Claude Code Remote Control	One session per connection	Yes	Yes	Yes	·	·
Claude Code agent view	Only after /bg	No, attach to answer	Queued	Whole session only	·	·
claude.ai/code	Cloud sessions only	·	Queued	·	·	·
nikitadoudikov/claude-pulse	Yes	Yes	·	·	·	·
bruceyxli/claude-code-monitor	Yes	Yes	·	·	·	·
tombelieber/claude-view	Yes	·	·	·	·	·
sverrirsig/claude-control	·	Yes	·	·	·	·
d-kimuson/claude-code-viewer	·	Yes	·	·	·	·
tiann/hapi	·	Yes	·	·	·	·
Justin0504/Aegis	·	Yes	·	·	·	·
mercurialsolo/claudectl	·	Its local model decides	Between its own agents	·	Prevents it in its own swarm	·
hahahahahahahahah6/edit-guard	·	·	·	·	Blocks stale edits	·
gantrol/AgentController	·	·	Codex only	·	·	·
Seven phone and chat approvers, listed below	·	Yes	·	·	·	·

Stars and dates in the lists below come from GitHub on 4 October 2026.

Official tools (5)

Tool	What it does
Claude Code agent view	One screen for your background Claude Code sessions, with claude agents
Claude Code Remote Control	Continues one local session from claude.ai/code or the Claude mobile app
Claude Code desktop app	Parallel sessions with git isolation, panes, a terminal, and a file editor
claude.ai/code	Claude Code sessions that run in the cloud
Codex app	The OpenAI client for parallel Codex agents

Dashboards and monitors (10)

Tool	Stars	What it does
disler/claude-code-hooks-multi-agent-observability	1,545	Shows Claude Code agents live through hook events
hoangsonww/Claude-Code-Agent-Monitor	1,039	Tracks Claude Code and Codex sessions in a dashboard
nikitadoudikov/claude-pulse	246	Watches every local Claude Code and Codex session, with approvals from a phone
tombelieber/claude-view	110	Live dashboard with cost tracking, search, and subagents
FlorianBruniaux/ccboard	96	Terminal and web dashboard for sessions, costs, and settings
bruceyxli/claude-code-monitor	13	Live dashboard for Claude Code sessions, with remote approval
szaher/claude-monitor	6	Shows token use, costs, and tool calls of your sessions
appaquet/ccmon	0	Live dashboard for Claude Code sessions
londondan/AgentConductor	0	Claude Code plugin that shows a live agent tree
Doogit/AgentWrangler	0	Shows token spend and session outcomes

Session managers that start the agents (26)

Tool	Stars	What it does
stablyai/orca	84,439	Runs many coding agents side by side, each in its own worktree
BloopAI/vibe-kanban	28,256	Kanban board for coding agents. Its README says that it is ending
manaflow-ai/cmux	27,599	macOS terminal with tabs and alerts for coding agents
winfunc/opcode	22,425	Desktop app and toolkit for Claude Code
superset-sh/superset	14,852	Agent workspace with terminals, code review, and previews
siteboon/claudecodeui	13,922	Web and mobile client for Claude Code, Codex, and others
Untrivial-ai/agent-orchestrator	12,688	Plans, runs, and supervises teams of coding agents
smtg-ai/claude-squad	8,566	Terminal app for many agents in separate workspaces
stravu/crystal	3,123	Parallel sessions in git worktrees. It is now Nimbalyst
omnara-ai/omnara	2,878	Platform to run and manage agents
nimbalyst/nimbalyst	1,828	Visual workspace for parallel coding agents
d-kimuson/claude-code-viewer	1,292	Full web client for Claude Code, with inline approvals
kbwo/ccmanager	1,257	Command line manager for sessions across worktrees
sugyan/claude-code-webui	1,137	Web chat for the Claude CLI. Archived
asheshgoplani/agent-deck	995	Terminal session manager for coding agents
Ark0N/Codeman	783	Self-hosted control center for agents in tmux
built-by-as/FleetCode	424	Runs command line agents in parallel worktrees
craftzdog/tmux-claude-hatch	395	tmux plugin with one Claude Code popup per project
imbue-ai/sculptor	235	Parallel coding agents in isolated containers
mercurialsolo/claudectl	202	Runs a swarm of agents with a local model that approves calls
sverrirsig/claude-control	135	macOS app that starts and manages sessions, with approvals
vultuk/claude-code-web	97	Web interface for the Claude Code CLI
OneStepAt4time/aegis	12	Starts and manages Claude Code sessions through an API
superkoh/koloft	2	macOS app that runs and manages sessions
StanislavBG/claude-code-session-manager	1	Schedules sessions and watches the token budget
Conductor	Closed source	Runs parallel agents in isolated workspaces on your Mac

Remote and mobile control (14)

Tool	Stars	What it does
slopus/happy	23,999	Mobile and web client for Claude Code and Codex
tiann/hapi	5,174	Controls agent sessions from a phone, the web, or Telegram
overwirehq/claude-code-telegram	2,798	Telegram bot for Claude Code
gbasin/agentboard	418	Web interface for tmux, made for agent terminals
onikan27/claude-code-monitor	310	Watches sessions from a terminal or a phone
yuuichieguchi/claude-remote-approver	74	Approves permission prompts from a phone
gantrol/AgentController	44	Controls Codex with a game controller
jsayubi/ccgram	27	Controls Claude Code from Telegram
coa00/claude-push	7	Sends permission requests to a phone through ntfy
nickknissen/claude-ntfy-hook	6	Allow or Deny from a phone through ntfy
tsyche/hookline	0	Approves permission prompts from a phone through ntfy
ssarnecki/claude-code-phone-notifications	0	Phone alerts with Approve, Approve All, and Deny
heywood8/claude-mobile-buddy	0	Approves prompts from an Android phone over Bluetooth
wolfpeter/claude-session-manager	0	Runs and watches sessions in tmux from a phone

Policy and safety tools (4)

Tool	Stars	What it does
Justin0504/Aegis	479	Policy rules, approvals, and a kill switch for agents
anipotts/cc	2	Tools for many Claude Code sessions. Archived
LoveCppp/AgentReins	0	Safety checks and recovery for coding agents on macOS
hahahahahahahahah6/edit-guard	0	A hook that blocks stale edits across sessions

Usage trackers such as ccusage, Claude-Code-Usage-Monitor, ccflare, and better-ccusage count tokens and costs. They do not control sessions, so they are not in the count.

If a tool is missing or a row is wrong, open an issue.

Security

AgentForeman can approve tool calls, so it guards its own page:

• The server listens on 127.0.0.1 only. Other computers cannot connect.
• The page and its data open only with a sign-in link. Other accounts on your Mac cannot use them.
• The link holds a token from ~/.agentforeman/token, which only you can read. The token stays the same across restarts.
• The browser keeps the token in a cookie that page scripts cannot read.
• Each button also sends the token. The server refuses a request with a different host name or from another web site.
• No other page can show AgentForeman inside a frame.
• AgentForeman trusts other programs of the same macOS user. They can read the same files.
• Rules are off by default, and they fail closed.

SECURITY.md has the full threat model and tells you how to report a problem.

How it works

1. Claude Code writes one small file for each live session in ~/.claude/sessions/.
2. AgentForeman keeps the sessions whose process is still alive.
3. For each session, it reads the end of its transcript in ~/.claude/projects/, then only new lines.
4. It reads the subagent files of each session, and the team files of Agent Teams.
5. For Codex, it reads the thread tables in ~/.codex in read-only mode.
6. About every 1.5 seconds, the server sends any change to the open page.

AgentForeman never writes to these folders. The controls work through small files in ~/.agentforeman, which only you can read. A hook writes a file when a tool call waits or a turn ends. The page shows it. When you click, the server writes your answer as a file, and the hook gives it to the session.

Each hook fails open, so a broken hook never blocks Claude Code. SPEC.md holds the full contract.

Requirements

• macOS. The tests do not cover Linux.
• Python 3.11 or later. Check with python3 --version.
• Claude Code.
• Codex is optional. The Codex page needs it.
• VS Code is optional. The button Open in VS Code needs it.

Fix common problems

• "Address already in use": AgentForeman already runs. Run agentforeman open, or use --port 4330.
• The page says "Open AgentForeman from your terminal": run agentforeman open.
• The sidebar says "Signed out": someone deleted ~/.agentforeman/token. Run agentforeman open.
• No sessions show: start a Claude Code session, then wait two seconds.
• Still no sessions: check that ~/.claude/sessions/ holds files.
• Open in VS Code fails: install VS Code in /Applications.
• Open in VS Code still fails: in VS Code, run "Shell Command: Install 'code' command in PATH".
• The button says Terminal: that session runs in a terminal, so switch to that terminal window.
• The sidebar says Controls off: run agentforeman install.
• The sidebar says Controls outdated: the hook changed in an update. Run agentforeman install again.
• A question shows no reply box: its turn ended before the install. Answer in the session this time.
• Approve does nothing: someone answered in VS Code, the terminal, or the Claude app first.
• A note never arrives: the agent makes no tool call now. The note goes away when the turn ends.
• A rule does not allow a call: check that rules are on.
• A rule still does not allow it: a rule never matches a command that it cannot read safely.
• Always allow is missing on a card: that rule grants too much, for example Bash(sudo apt).
• Alerts stay off: allow notifications for 127.0.0.1 in your browser, then click Alerts off again.
• The first load is slow: security software on some Macs scans each file that AgentForeman opens.
• A hook seems to do nothing: create the file ~/.agentforeman/debug, then read ~/.agentforeman/hook.log.

Development

Every change starts in SPEC.md. CONTRIBUTING.md has the details.

python3 -m venv .venv
. .venv/bin/activate
pip install -e ".[dev]"
python -m playwright install chromium
ruff check .
ruff format --check .
pytest -q

To record the GIFs again, install ffmpeg and run python tools/record_gifs.py. To record the one-minute tour, also install Pillow and run python tools/record_showcase.py.

License

MIT. See LICENSE.