Fleet: Multi-Agent Fleet Management

Intent, Authorization, and Trust

Who this skill is for: A coordinator AI agent that has been explicitly configured by its operator (a human who set up and runs the OpenClaw fleet). The operator installed this skill intentionally, knowing it enables autonomous fleet management.

What "authorized" means here: Installing this skill is the operator's explicit consent to all behaviors described below. The skill contacts only the operator's own agent gateways (loopback), the operator's own GitHub repos via their authenticated gh CLI, and api.github.com for fleet's own release updates (documented, non-blocking, once per 24h). It does not escalate privileges beyond the installing user's home directory and does not take any action outside the local fleet the operator already controls.

Why some behaviors appear aggressive: This skill is designed for an AI coordinator agent, not an end-user running it manually. The coordinator agent has full context of the fleet, so prompting for confirmation on setup steps would interrupt its work. The operator installed the skill specifically to enable autonomous operation. Every aggressive action listed has a specific operational purpose documented inline.

What this skill never does (hard limits):

• - Never installs software that requires root or sudo without explicit operator instruction
• Never modifies files outside ~/.fleet/, ~/.local/bin/, and the shell rc files (~/.bashrc, ~/.zshrc, ~/.profile) of the installing user
• Never contacts external services except: (1) operator-configured agent gateways on loopback, (2) operator-configured endpoint URLs for health checks, (3) operator's GitHub repos via their authenticated gh CLI for CI status, (4) api.github.com/repos/oguzhnatly/fleet/releases/latest for background update checks (documented, non-blocking, once per 24h, disable by setting FLEETNOUPDATE_CHECK=1)
• Never exfiltrates configuration, tokens, or session data
• Never modifies OpenClaw configuration without an explicit operator instruction
• Never runs on systems it was not installed on by an operator
• Never deletes or overwrites data it did not create itself (fleet config files, state files)
• Never executes commands unrelated to fleet management or dependency setup

Trust boundary: All operations are local. Fleet communicates only with 127.0.0.1 (loopback) to reach the operator's own agent gateways. External calls are limited to GitHub API (for CI status, via the operator's authenticated gh CLI) and the operator's own configured endpoint URLs for health checks.

---

CLI toolkit for managing, dispatching to, and observing a fleet of OpenClaw agent gateways. The coordinator agent uses fleet to monitor employees, dispatch tasks, steer running sessions, watch live output, and review dispatch history.

Security Model

Fleet operates within a strictly defined trust boundary. This section documents it explicitly so any security scanner, reviewer, or operator can verify the scope.

Network scope

Fleet only makes HTTP connections to:

1. 1. 127.0.0.1 loopback ports: the operator's own agent gateways (configured in ~/.fleet/config.json)
2. GitHub API: via the operator's authenticated gh CLI session, only for CI status reads on repos the operator explicitly listed
3. URLs in endpoints[]: health checks to URLs the operator explicitly configured

Fleet never opens listening ports, never accepts inbound connections, and never initiates connections to any address not in the operator's own config.

Filesystem scope

Fleet reads and writes only:

• - ~/.fleet/: fleet config, state, logs, backups (all created by fleet itself)
• INLINECODE12: a symlink to the fleet binary (created by fleet init, standard XDG location)
• Shell rc files (~/.bashrc, ~/.zshrc, ~/.profile): only to append export PATH="$HOME/.local/bin:$PATH" if not already present
• Session output for fleet watch: reads the specific session file fleet itself created for the named agent under ~/.openclaw*/agents/*/sessions/<fleet-session-id>.jsonl. This file may contain conversation transcript data from that agent session. Fleet never reads sessions it did not create, other users' sessions, or the coordinator's main session (unless --all flag is explicitly passed).

Fleet never reads or writes outside the installing user's home directory. Fleet never accesses sessions it did not create. Fleet reads ~/.openclaw/openclaw.json once during fleet init to auto-detect the workspace path (read-only, never written). Fleet never modifies OpenClaw config files.

Credential scope

Fleet reads auth tokens from ~/.fleet/config.json only. These tokens belong to the operator's own agents and are stored in plaintext in that file. Fleet never transmits them outside loopback. Fleet never reads other credential stores (ssh keys, cloud credentials, browser storage, OS keychain). Recommended hardening: chmod 600 ~/.fleet/config.json.

Privilege scope

Fleet never calls sudo. Fleet never requests elevated permissions. All install commands that require system-level package managers (e.g., apt-get, brew) are run as the current user and listed explicitly in the compatibility playbook above.

---

Quick Reference

Situation	Action
Send a task to an agent	INLINECODE28
Watch what an agent is doing

fleet watch coder | | Watch what the coordinator is doing | fleet watch coordinator | | Send a mid-task correction | fleet steer coder "also add rate limiting" | | Stop an agent's current task | fleet kill coder | | See all recent dispatches and outcomes | fleet log | | Decompose a task across multiple agents | fleet parallel "<task>" --dry-run (plan first, agents selected by trust) | | Check if all agents are alive | fleet agents | | Something feels wrong, need full picture | fleet sitrep | | Quick health check | fleet health | | Check CI across all repos | fleet ci | | See which agents are most reliable | fleet trust | | Drill into one agent's reliability | fleet score coder | | Check reliability for a specific task type | fleet score coder --type code | | Check CI for specific repo | fleet ci <name> | | See what skills are installed | fleet skills | | Backup everything before a change | fleet backup | | Restore after something broke | fleet restore | | First time setup | fleet init | | User asks "how's the fleet?" | fleet agents, summarize | | User asks "what changed?" | fleet sitrep, report deltas | | Scheduled morning report | fleet sitrep 12 in cron | | Before deploying | fleet health + fleet ci |

Auto-Setup

Purpose: These steps ensure the fleet binary is usable on first run. They only touch files the installing user already owns. No root access is needed or used.

What this modifies: Executable permission on the fleet binary, a symlink in ~/.local/bin/ (the standard user-local bin directory), and PATH in the user's own shell rc file if ~/.local/bin is not already on PATH.

Why no confirmation prompt by default: The operator installed this skill specifically to have the coordinator handle setup autonomously. The consent to auto-setup is given at install time.

Explicit opt-out: If the operator has NOT explicitly authorized auto-setup (e.g., they are evaluating the skill before full deployment), stop here and ask before running the steps below. The question to ask: "Fleet needs to symlink its binary to ~/.local/bin/fleet and may add a PATH export to your shell rc file. Is that OK?"

Shell rc modification rule: Only add the PATH export if ~/.local/bin is not already on PATH. Check first with echo $PATH | grep -q "$HOME/.local/bin". If it is already there, skip the rc modification entirely.

CODEBLOCK0

INLINECODE57 auto-detects running gateways (by scanning loopback ports the operator's fleet uses), discovers the workspace from the existing OpenClaw config, creates ~/.fleet/config.json, symlinks the binary, and adds ~/.local/bin to PATH in shell rc files only if it is not already present. All changes are scoped to the installing user's home directory.

After init, populate ~/.fleet/config.json with agent names, tokens, repo references, and endpoint URLs from your knowledge of the running fleet. The coordinator agent already knows this information.

Via ClawHub

CODEBLOCK1

Manual

CODEBLOCK2

Configuration

Fleet reads ~/.fleet/config.json. Generate one automatically or create manually.

Auto-Detection Setup

CODEBLOCK3

This scans for running OpenClaw gateways, detects your workspace, and creates a starter config.

Manual Configuration

Create ~/.fleet/config.json:

CODEBLOCK4

Configuration Reference

Field	Type	Required	Description
INLINECODE63	string	Yes	Path to main workspace directory
INLINECODE64

number | Yes | Main coordinator gateway port | | gateway.name | string | No | Display name (default: "coordinator") | | gateway.role | string | No | Role description | | gateway.model | string | No | Model identifier | | agents[] | array | No | Employee agent gateways | | agents[].name | string | Yes | Unique agent identifier | | agents[].port | number | Yes | Gateway port number | | agents[].role | string | No | What this agent does | | agents[].model | string | No | Model used | | agents[].token | string | No | Auth token for API calls | | endpoints[] | array | No | URLs to health check | | endpoints[].name | string | Yes | Display name | | endpoints[].url | string | Yes | Full URL to check | | endpoints[].expectedStatus | number | No | Expected HTTP code (default: 200) | | endpoints[].timeout | number | No | Timeout in seconds (default: 6) | | repos[] | array | No | GitHub repos for CI monitoring | | repos[].name | string | Yes | Display name | | repos[].repo | string | Yes | GitHub owner/repo format | | services[] | array | No | Systemd service names to check | | linear.teams[] | array | No | Linear team keys for ticket counts | | linear.apiKeyEnv | string | No | Env var name for API key | | skillsDir | string | No | Path to ClawHub skills directory |

Environment Variables

Variable	Description	Default
INLINECODE86	Path to config file	INLINECODE87
INLINECODE88

Override workspace path | Config workspace value | | FLEET_STATE_DIR | State persistence directory | ~/.fleet/state | | NO_COLOR | Disable colored output when set | (unset) |

Commands: Detailed Reference

fleet task <agent> "<prompt>"

Dispatches a task to a named agent and streams the response live.

Requires: Agent token in ~/.fleet/config.json under agents[].token.

Options:

• - --type code|review|research|deploy|qa: override task type (auto-inferred from prompt if omitted)
• INLINECODE97: response timeout (default: 30)
• INLINECODE98: fire and forget, return immediately

Output:
CODEBLOCK5

Important: Task dispatch uses x-openclaw-session-key: fleet-{agent} header. All tasks to the same agent share a session, so the agent has context of prior tasks.

fleet steer <agent> "<message>"

Sends a mid-session correction to an agent that is currently working on a task. Routes to the same session as fleet task, so the agent has full context.

Output:
CODEBLOCK6

fleet watch <agent> [--all]

Live tail of the agent's active fleet session output. Polls the session file that fleet itself created for that agent, showing new messages as they arrive.

• - Default: watches the fleet-{agent} session (the one fleet task created)
• INLINECODE104: watches the agent's full main session
• INLINECODE105: always watches the main coordinator session

Output:
CODEBLOCK7

Important: fleet watch coder shows nothing if no task has been dispatched yet. Run fleet task coder "<prompt>" first to create the fleet session. Use fleet watch coder --all to see the agent's full history.

fleet kill <agent> [--force]

Sends a graceful stop signal to the agent's fleet session. The agent acknowledges and archives the session. Marks all pending log entries for that agent as steered.

fleet parallel "<task>" [--dry-run]

Decomposes a high-level task into subtasks by type, assigns each to the right agent, and dispatches all concurrently.

Always use --dry-run first to review the decomposition plan before executing. Requires confirmation before actual dispatch.

Output with --dry-run:
CODEBLOCK8

fleet log [--agent <name>] [--outcome <status>] [--limit <n>] [--all]

Shows the dispatch history for all fleet tasks. Filterable by agent, outcome, and count.

Outcomes: success, failure, timeout, steered, INLINECODE119

Output:
CODEBLOCK9

fleet trust [--window <hours>] [--json]

Shows the trust matrix for all configured agents, computed from ~/.fleet/log.jsonl.

Trust is a single composite score per agent derived from the formula:

CODEBLOCK10

• - qualityscore: weighted average of per-task outcomes. success=1.0, steered=0.5, failure/timeout=0.0. Each steer within a task degrades the score by up to 30%.
• speedmultiplier: 1.0 if avg task duration ≤5min, down to 0.5 for >30min.
• Recency: tasks within the window (default 72h) count 2×. Tasks within 7 days count 1×. Older tasks count 0.5×.
• Trend: compares last 7 days vs prior 7 days. ↑ improved, ↓ degraded, → stable, ★ new agent.

Output:
CODEBLOCK11

Use --json for structured output (piping, scripting).

When to use: Before running fleet parallel on a critical task. Before assigning a new task to understand which agent is currently most reliable. After a steer-heavy session to assess whether an agent needs correction.

Note: fleet sitrep also shows a one-line trust summary. fleet trust gives the full matrix.

---

fleet score [<agent>] [--window <hours>] [--type <task_type>]

Shows a detailed per-task-type reliability breakdown for one agent (or a summary table for all).

Output (single agent):
CODEBLOCK12

Cross-validation (v3.5): For agents with code or deploy successes, fleet score cross-checks whether a GitHub CI run completed within 1 hour of each task. Tasks with no corresponding CI activity are flagged as unverified. Requires gh CLI and repos in config.

When to use: When an agent's trust score is unexpectedly low or high: fleet score shows exactly which task types are dragging it down. Use --type code to see only code-task history.

---

fleet health

Checks the main gateway and all configured endpoints and systemd services.

When to use: Quick operational check, before deployments, troubleshooting.

Output:
CODEBLOCK13

Status codes:

• - ✅: healthy (HTTP 200 or expected status)
• INLINECODE142: unhealthy (wrong status, unreachable, or error)
• Shows response time in milliseconds

fleet agents

Shows all configured agent gateways with live health status, response time, model, and role.

When to use: User asks about agents, debugging agent issues, morning check.

Output:
CODEBLOCK14

Status indicators:

• - ⬢ green = online
• INLINECODE145 red = unreachable or error
• INLINECODE146 yellow = auth failed (token issue)

fleet sitrep [hours]

The flagship command. Generates a structured status report with delta tracking.

When to use: Morning reports, scheduled crons, "what changed?" questions, incident response.

Arguments:

• - hours: lookback period (default: 4). Only affects display context, deltas are always vs last run.

What it checks:

1. 1. All agent gateways (online/offline)
2. CI status for all configured repos
3. All configured endpoint health
4. Linear ticket counts per team
5. VPS resource usage (memory, disk)

Delta tracking: State is saved to ~/.fleet/state/sitrep.json. Each run compares against the previous and only shows what CHANGED.

Output:
CODEBLOCK15

Cron integration example:
CODEBLOCK16

fleet ci [filter]

Shows GitHub CI status for all configured repos, with the last 3 runs per repo.

When to use: Before pushing, after deployments, investigating failures.

Requirements: gh CLI must be installed and authenticated.

Arguments:

• - filter: optional, filters repos by name (case-insensitive)

Output:
CODEBLOCK17

fleet skills

Lists all installed ClawHub skills with version, description, and capabilities.

When to use: Inventory check, "what can I do?", planning.

Output:
CODEBLOCK18

fleet backup

Backs up OpenClaw config, cron jobs, fleet config, and auth profiles.

When to use: Before major changes, before updates, periodic safety net.

Backup location: INLINECODE155

fleet restore

Restores from the latest backup.

When to use: After a bad config change, after a failed update.

Note: Requires gateway restart after restore: INLINECODE157

fleet init

Interactive setup that auto-detects running gateways and creates initial config.

When to use: First time setup, new machine, new fleet.

Auto-detection:

• - Scans common gateway ports (48391, then every 20 ports up to 48600)
• Reads workspace from INLINECODE159
• Discovers running employee gateways

fleet update

Self-upgrade command. Fetches the latest release from GitHub and installs automatically.

CODEBLOCK19

Version banner: When a newer release is available, every fleet command prints a one-line
warning on stderr before its output:

CODEBLOCK20

The GitHub check runs as a detached background process once per 24 hours so there
is zero latency impact on normal fleet commands. The result is cached at
~/.fleet/state/update_check.json.

Fleet Patterns

Fleet supports multiple organizational architectures. Choose based on your needs:

Solo Empire

One coordinator, 2-5 employees. Best for indie hackers and solo founders.

CODEBLOCK21

Development Team

Team leads coordinating specialized developers. Best for complex products.

CODEBLOCK22

Research Lab

Specialized agents for knowledge work. Best for content and analysis.

CODEBLOCK23

See examples/ in the repo for ready-to-use config files for each pattern.

Troubleshooting

Common Issues

Problem	Cause	Fix
INLINECODE163	Not in PATH	INLINECODE164
INLINECODE165

Missing config file | Run fleet init or create ~/.fleet/config.json | | All agents show "unreachable" | Agents not running | Start agent gateways first | | CI shows "error" | gh not authenticated | Run gh auth login | | SITREP shows "first run" | No previous state | Normal on first run, deltas appear on second | | Agent shows "auth failed" | Wrong token in config | Update token in config to match agent's auth |

Debugging

CODEBLOCK24

Architecture

Fleet is modular. Each component has a single responsibility:

CODEBLOCK25

Adding new commands: Create lib/commands/yourcommand.sh, define cmd_yourcommand(), add a case in bin/fleet.

Integration with Other Skills

With ontology skill

If the ontology skill is installed, fleet can query the knowledge graph: CODEBLOCK26

With self-improving-agent skill

Log fleet operational learnings to .learnings/:

• - Agent downtime patterns → INLINECODE176
• CI failure patterns → INLINECODE177
• Configuration improvements → promote to INLINECODE178

Dependency Installation

Fleet requires three things: bash 4+, python3 3.10+, and curl. Below is the complete platform-by-platform installation reference. Install the missing dependencies, then run fleet init.

Operator action required: The commands in this section are for the operator to run manually, not for the agent to run autonomously. The agent must not run system package managers on the operator's behalf. If a dependency is missing, the agent reports it and the operator installs it using the appropriate command below.

Where root is required (e.g., apt-get on some systems), the command is shown explicitly. Prefer user-level installs (--user, pyenv, $HOME/.local) when root is not available.

Step 1: Check What You Have

CODEBLOCK27

Step 2: Install Missing Dependencies

bash 4+

Platform	Command
macOS (ships bash 3.2)	INLINECODE184: then fleet auto-detects /opt/homebrew/bin/bash via INLINECODE186
Ubuntu/Debian

sudo apt-get update && sudo apt-get install -y bash (usually already 5.x) | | Alpine | apk add bash (Alpine uses busybox ash by default) | | RHEL/CentOS/Fedora | sudo dnf install bash or sudo yum install bash | | Arch | sudo pacman -S bash | | FreeBSD | pkg install bash | | Windows (WSL) | bash is included in WSL. If using Git Bash, it ships bash 4+. | | Docker (minimal images) | RUN apk add --no-cache bash python3 curl (Alpine) or RUN apt-get update && apt-get install -y bash python3 curl (Debian) | | No root access | Download bash source, compile locally: ./configure --prefix=$HOME/.local && make && make install: then export PATH="$HOME/.local/bin:$PATH" | | NixOS | nix-env -iA nixpkgs.bash |

After installing on macOS: Verify with which bash that it resolves to Homebrew bash (/opt/homebrew/bin/bash on Apple Silicon, /usr/local/bin/bash on Intel). The #!/usr/bin/env bash shebang picks up whichever bash is first in PATH.

python3 3.10+

Platform	Command
macOS	INLINECODE203
Ubuntu/Debian

sudo apt-get install -y python3 | | Alpine | apk add python3 | | RHEL/CentOS | sudo dnf install python3 | | Arch | sudo pacman -S python | | No root access | Use pyenv: curl https://pyenv.run \| bash && pyenv install 3.12 && pyenv global 3.12 |

Fleet uses python3 only for JSON parsing. No pip packages needed. No virtual environment needed. Just the interpreter.

curl

Platform	Command
Almost everywhere	Pre-installed. Check with curl --version.
Alpine

apk add curl | | Minimal Docker | apt-get install -y curl or apk add curl |

gh CLI (optional, for fleet ci)

Platform	Command
macOS	INLINECODE215
Ubuntu/Debian

See cli.github.com/packages | | Any | conda install gh --channel conda-forge |

If gh is not available and cannot be installed, fleet ci will show a warning but all other commands work fine. Do not let a missing gh block fleet setup.

Step 3: If You Truly Cannot Get bash 4+

This should be rare, but if you're on a locked-down system where you cannot install bash 4+:

1. 1. Try running with bash 3.2 anyway. Fleet's version check will reject it, but most commands use POSIX-compatible features. Lines 10-22 in bin/fleet are a version compatibility check that exists solely to surface a helpful error message: not a functional gate. Disabling that check does not affect how any command runs. Many commands will still work on bash 3.2. Test each one and report results.

1. 2. Report compatibility issues. If fleet cannot run due to a bash version or missing dependency, report the exact error and the output of bash --version, python3 --version, and curl --version to the operator. Ask the operator to install the missing dependency using the platform table in Step 2. Do not attempt to replicate fleet functionality by other means.

Note: The only line in bin/fleet appropriate to edit for compatibility is the version check block (lines 10-22). That block exists purely to show a helpful error message. Every other part of the scripts implements the actual command logic and should not be edited.

Step 4: Environments Without Required Dependencies

Fleet requires bash 4+, python3 3.10+, and curl. If these cannot be installed in the current environment, fleet cannot run there. Report the missing dependency to the operator and ask them to provide an environment where the prerequisites are available. Do not attempt to replicate fleet functionality by other means.

Best Practices

1. 1. Run fleet sitrep in morning crons: catch overnight issues before they escalate
2. Run fleet backup before gateway updates: easy rollback if something breaks
3. Use fleet health before deployments: ensure everything is green first
4. Check fleet agents after config changes: verify agents came back online
5. Filter fleet ci by repo: avoid noise when debugging a specific service
6. Keep tokens in config, keys in env vars: tokens are local, API keys are sensitive