vitals

Give your AI agent a memory of its own infrastructure.

Quick start

uv tool install agent-vitals
av install    # detects hosts, wires MCP + priming skill

That's it. 30 seconds, no daemon, no cloud, no accounts.

After install, restart your agent host so it picks up the new MCP server. Then the next time it runs, it'll call vitals_summary before risky operations — automatically.

• Shadow & stale detection — cron, systemd, MCP, skills. Shows everything; filters to broken references only.

• Burnout tracking — per-agent completion rates, trends, stuck-session heuristics.

• Doom-loop detection — exact + soft loops in Bash and Edit. Excludes polling; compares edits by content.

• Unused-tool finder — per-server + per-tool usage. GitHub-measured 8–12KB overhead per unused tool, per turn.

• Token cost + Effective Tokens — model-aware pricing, ET metric, downgrade suggestions.

• SSH polling detector — finds repeated SSH commands to the same host.

• Overlap detection — duplicate/similar tool names across MCP servers.

• Session compaction — keep last N events, backup first, dry-run preview.

• Coaching — generates optimized system prompts from your actual session data.

• Trace suite — recorded session forensics for Claude Code and pi. av trace list/summary/replay/diff/errors/profile/grep/export/watch/suggest. No payloads, no secrets.

• Pre-action hooks — PATH wrappers around crontab/systemctl that gate mutations until vitals is fresh.

Related MCP server: iris-eval/mcp-server

What your agent sees

When your agent (pi, Claude Code, OpenCode, Cursor, Codex) calls vitals_summary, it gets:

  ▄▀ vitals — local agent stack health

  shadow: 13 autonomous thing(s) configured (mcp: 8, systemd: 5)
  subagent burnout (7d): 5 runs, 100% completion ✓
  claude code (7d): 170 sessions, 45,626 events, ⚠ 36 stuck
    - biggest stuck-looking session: 7,118 events

That's not a CLI you read. That's a tool your agent calls from inside its own loop — before it does anything risky.

How it works

Three moves. No daemon. No cloud.

1. Install — uv tool install agent-vitals && av install. Registers the MCP server and priming skill for your agent host.

2. Observe — the agent calls vitals_summary before risky operations. Hooks gate crontab / systemctl until vitals is fresh.

3. Optimize — run av cost, av loops, av coach to extract patterns and shrink the model's workload.

Before / after

User: schedule a daily backup at 2am

Agent: yep, done.

  0 2 * * * /home/me/backup.sh
        >> /tmp/backup.log 2>&1

… and /home/me/backup.sh doesn't exist. Crontab fails silently forever.

User: schedule a daily backup at 2am

Agent: one moment — verifying first.

  → shadow_stale       (no other cron broken)
  → check backup.sh    (exists? writable?)
  → write crontab entry
  → shadow_stale       (confirm no breakage)

  ✓ crontab updated.

User: is my agent setup healthy?

Agent: looks fine to me!

(blind vibes)

User: is my agent setup healthy?

Agent: → vitals_summary

  shadow: 13 autonomous thing(s) configured
  claude code (7d): ⚠ 36 stuck sessions
    - biggest: 7,118 events

  ⚠ 36 sessions look stuck. want me to triage?

The diff is vibes vs. data.

Install

uv tool install agent-vitals
av install  # detects every agent host on your box and wires them up
            # restart your agent host so it picks up the new MCP server

That's the whole setup. 30 seconds.

Pre-action hooks (v0.3.0)

Priming asks the agent to call vitals. Hooks refuse the operation when it hasn't.

av hooks install     # one-time setup (~3 seconds)
av hooks status      # check freshness
av hooks disable     # temporarily turn off (rename to *.disabled)
av hooks uninstall   # full removal

After av hooks install, the one-liner above is appended to your ~/.bashrc / ~/.zshrc. Open a new terminal and crontab -e or systemctl --user enable foo will be refused at the OS level unless an agent (or av doctor) refreshed the vitals stamp in the last 60 seconds.

  ⚡ vitals hook: refused `crontab -e`

  reason:   vitals stamp is 5m12s old — exceeds 60s window.
  stamp:    5m12s ago

  refresh:  call any vitals tool or run `av doctor`
  bypass:   VITALS_BYPASS=1 crontab -e

What's gated: crontab -e/-r/-i/<file>/- and systemctl --user {enable,disable,start,stop,restart,reload,mask,unmask,daemon-reload,...}.

What's NOT gated: crontab -l (reads), systemctl status / list-* / is-active / show / cat (reads), and power management (reboot, poweroff, suspend) — a stale stamp must never block a reboot.

Bypass for emergencies: VITALS_BYPASS=1 crontab -e skips the check.

What av install does

  $ av install
                    detected 3 agent host(s)
  ┏━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┓
  ┃ host        ┃ config                           ┃ status   ┃
  ┡━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━┩
  │ pi          │ ~/.pi/agent/mcp.json │ detected │
  │ Claude Code │ ~/.claude/.mcp.json  │ detected │
  │ Codex CLI   │ ~/.codex/config.toml │ detected │
  └─────────────┴──────────────────────────────────┴────────━━┘

  installing:

  ┏━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┓
  ┃ host        ┃ mcp config         ┃ skill/rule        ┃
  ┡━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━┩
  │ pi          │ added              │ installed         ┃
  │ Claude Code │ added              │ already installed ┃
  │ Codex CLI   │ added              │ installed         ┃
  └─────────────┴────────────────────┴────────────────━━━┛

  ✓ done. Restart your agent host.

The five tools

All tools are local-only, read-only, safe to call repeatedly. None of them modify state.

The trigger table

The table av install installs into your priming skill — so the agent knows when to reach for each tool without you asking:

v0.3.0 changes this. av hooks install deploys PATH-level wrappers around crontab and systemctl --user that refuse any mutation when the vitals stamp is older than 60 seconds. Read operations (crontab -l, systemctl status, etc.) are never gated. See Pre-action hooks below.

Anti-patterns this exists to prevent

• ❌ "Your crontab is fine" — without calling shadow_stale first

• ❌ Scheduling cron / systemd work — without verifying the target binary exists

• ❌ Starting a 4-hour task — while 6 other sessions are stuck on the same box

• ❌ Pretending a task completed — without checking burnout_summary

• ❌ Recommending an MCP install — without shadow_list to check for duplicates

• ❌ Debugging slowness — without first checking vitals_summary

What it scans

av                # one-shot health summary
av doctor         # summary + actionable recommendations
av shadow         # what's configured on your box
av shadow --watch # live refresh every 2s
av burnout        # completion metrics, last 7 days
av burnout --days 30
av detect         # list detected agent hosts
av install        # interactive installer — pick components, pick hosts
av init           # deprecated alias for `av install`
av drift          # find inconsistencies across hosts (MCP, skills, hooks)
av sessions       # list session files with age + size
av snapshot       # tar.gz of agent state (mcp configs, skills, hooks)
av mcp            # start the MCP server (stdio)

# Efficiency suite (v0.6.0+)
av loops          # detect exact + soft doom loops in sessions
av loops -n 50     # show up to 50 findings
av unused         # find unused MCP servers + per-tool usage
av cost           # token spend with model-aware pricing + Effective Tokens
av tokens         # identify token-heavy tools
av tokens --suggest  # get optimization suggestions
av ssh            # detect SSH polling loops
av overlap        # detect overlapping MCP tools
av compact        # compact large session files
av compact --dry-run  # preview compaction without changes
av coach          # generate optimized system prompts from session data

# Trace (v0.7.0)
av trace list                  # list sessions with event counts and source type
av trace summary <session>     # turns, tools, errors, wall duration
av trace replay <session>      # step-by-step replay (no payloads)
av trace diff <a> <b>          # structural diff between two sessions
av trace errors <session>      # show only error events
av trace profile <session>     # per-tool breakdown: calls, errors, avg duration
av trace grep <session> <pat>  # filter events by tool name or type
av trace export <session>      # export normalized events to JSON
av trace suggest <session>     # actionable suggestions from session data
av trace watch <session>       # tail a session JSONL live (Ctrl-C to stop)
uv              ──  one-tool install / build / publish
typer           ──  CLI
rich            ──  terminal rendering
pyyaml          ──  SKILL.md frontmatter parsing
mcp (FastMCP)   ──  MCP server, stdio transport
pytest          ──  295 tests across 12 modules

~6000 lines of Python + 295 tests + priming SKILL.md. GPL v3 licensed.

Roadmap

• v0.1.0 — shadow + burnout CLI commands

• v0.2.0 — MCP server + av install for 5 host types

• v0.7.0 — trace suite: av trace list/summary/replay/diff/errors/profile/grep/export/watch/suggest. Content-agnostic adapters for Claude Code and pi JSONLs. MCP tools exposed.

• v0.7.0 — shadow live (running agent processes, ps-tree view)

• v0.8.0 — cross-session "agent déjà vu" detector (you researched this codebase 3 weeks ago)

Contributing

Issues and PRs welcome. Two things to know:

1. Scanners in src/agent_vitals/scanners.py are independent and fail gracefully. Add a new source by writing one scan_*() function and adding it to scan_all().

2. MCP tool docstrings are the product. The docstring on vitals_summary is the instruction the agent reads. Write it as a directive to the agent ("always call this first when…"), not API docs.

When you open a PR, paste the output of av shadow on your box so we can see what surfaces in your environment.

License

GPL v3. See LICENSE.