Guide 53 of 61 · Running a fleet with Central Lite
Roll out outbound inspection across a fleet
BetaList the sites your fleet's agents should not send secrets to, run them in Audit, read what Block would stop, then block through a canary rollout.
Built, not yet claimed preview: this guide describes a package that has not passed its release gate.
Choose the paste sites, file-sharing sites, webhooks and sites of your own that the fleet's launched agents should not quietly send secrets to, roll the list out in Audit from your policy repository, read Central Lite's replay of what Block would have stopped, and switch to Block through a canary that pauses by itself when machines report trouble. At the end you also know how to handle the tools that refuse AgentGarm's certificate and what to tell the people the list covers.
1. Choose presets and your own sites
In the console, Outbound inspection sits in the Policy group, for owners, admins, security and auditors (never finops or viewer). With nothing listed it offers Choose sites, four steps: what it does, the presets, starting in Audit, and deciding in 7 days. Edit sites does the same on a list that exists. The presets are lists AgentGarm keeps and checks against DNS at every catalog release, each shown with its count and the day it was checked: paste sites, file sharing and webhooks. Your own sites go one per line: an exact name, *. for every subdomain (not the name itself), or :port for another port than 443. The sheet converts a name with other letters to its xn-- form and refuses, as you type, an address, a single label, a public suffix, a model provider's or subscription host and a never-inspected host, in one sentence each; a *. entry on a shared domain such as ngrok-free.app is allowed with a note that it covers every tenant.
Outbound inspection is a public beta, with the exit criteria and the feedback route in the single-machine guide. The console's section has Send feedback too: a report of counts, versions and machine reasons only.
inspect:
presets: [paste_sites, file_sharing, webhooks]
# Your own sites: an exact name, *. for every subdomain (not the name itself), :port for
# another port than 443. Lowercase A-labels.
hosts: [uploads.halden.example, "*.hooks.halden.example", "files.halden.example:8443"]See also
Check what agents upload to paste sites and webhooksHave AgentGarm read what launched agents send to paste sites, file sharing and webhooks, record or stop secrets, and fix tools that refuse its certificate.
2. Start in Audit
“New sites start in Audit”: requests are sent as before and each machine records what Block would have stopped, so nobody's work is interrupted while you learn what the agents send. Alert sends the request and raises an alert, and Block stops it, both as soon as the change rolls out. A policy in audit mode keeps every site in Audit whatever mode says.
Central Lite pulls policy from your git repository and has no drafts, so Copy policy change and preview copies the network.inspect section of agents.d/default.yaml as a diff, with the commit, garm-central policy attest … --write-note and push commands, and its reach (how many policies and machines; agents already running keep their list until restarted). Each task follows the section of its own policy, so give an agent's own policy file the section too if it should inspect. Central Lite's lint gate decides at publication, and the rollout runs as any other: replay, canary, a quarter of the fleet, then every machine. A machine whose agentgarmd is too old for the section keeps its previous bundle, and the preview names it.
# New sites start in Audit.
mode: auditgarm-central policy attest --repo policies --source-repo "$PWD/policies" --dir committer --write-note
garm-central policy publish
garm-central rollout listSee also
Roll out policy from git with canariesKeep agents.d policies in git and let Central Lite roll each signed commit out to canaries, then the fleet, halting by itself when machines report trouble.
3. Read the replay
The editor's preview, and the section once sites are in Audit, say what Block would change, from what Central Lite has. Before Audit it has no record of these connections by site: they are encrypted end to end, and machines upload hosts only as keyed hashes. It counts connections only from recorded days you replay: agentgarmd replay export writes one user's day for one policy on a machine (refused under the non-monitoring profile), and replaying it against the commit or rollout with garm-central policy replay or rollout replay lets Central Lite count its connections per listed site, shown as “uploaded node-days”. It keeps those counts per recorded day and site for 90 days, never the day itself, so the same day replayed twice counts once; an export names neither machine nor user, so the machines it shows are recorded days. Which requests carry secrets can't be known before Audit, and the preview says so.
After Audit (7 days is the suggestion), the machines' ten-minute counts, the findings Block would have stopped and the tools that refused the certificate show per site, “on at least N machines”, with Switch to Block with a canary and Stay in Audit. The site table shows each site's mode, requests over 7 days with a trend, secrets, requests sent uninspected and the top agent; sites no policy lists any more stay as “No longer listed” while they have activity.
sudo agentgarmd replay export --policy default --uid 1000 --day 2026-11-10 --out default-day.jsonlgarm-central policy replay --commit --events default=default-day.jsonl --home /home/alice4. Switch to Block with a canary
Switch to Block with a canary opens the editor in Block; or set one site's mode under per_host, as here. The policy needs mode: enforce. Commit, attest and publish: the rollout reaches your canaries first, and besides its usual checks, four of outbound inspection's pause it by themselves, over the machines it reached: handshakes refused by a tool at 1 % or more of connections (judged from 100), any machine whose web CA is not ready, 5 or more stops per machine-day, and an added latency p99 of 8 ms or more. garm-central rollout config sets each threshold for new rollouts.
Two limits of the beta: what a machine reports as connections is requests inspected plus refused handshakes, so the 1 % is of requests and looser than a share of connections, and one request with several stopped findings counts more than once; and machines do not measure the added latency yet (they report 0), so that check cannot pause a rollout today. A paused rollout resumes, or rolls back, as in the policy rollout guide.
per_host:
# Block on one site first, through a canary.
paste.rs:
mode: blockgarm-central rollout config --inspect-refused-basis-points 100 --inspect-ca-not-ready-max 0 --inspect-stopped-per-day 5 --inspect-latency-p99-us 8000garm-central rollout resume <rollout id>5. Handle fallbacks
Above the site table the section pins the decisions it needs from you. “A tool refuses AgentGarm's certificate on N machines”: in Audit and Alert those requests were sent uninspected, in Block they are stopped. Exempt this site opens the editor with the site under exempt (one of your own sites is taken off the list instead), Keep stopping hides the row until more machines refuse, and See the machines counts them: Central Lite keeps host names and executables as keyed hashes, so the machine's own dashboard names the tool. A tool can often be made to trust the certificate (each runtime, Java, Go on macOS).
“A site asks for a client certificate”: inspection can't present one, so those requests go as a plain tunnel and are recorded; on_uninspectable: refuse refuses them instead. An approved exception is ignore_rules: that rule's findings on that site are recorded as expected. Under Exceptions and advanced the editor also sets WebSockets on listed sites (passed through uninspected and said so, or refused) and uploads over 8 MB (inspected as they stream, or refused in Block).
# This site asks for a client certificate: refuse it rather than tunnel it.
"files.halden.example:8443":
on_uninspectable: refuse
# An approved exception: this rule's findings here are recorded as expected.
uploads.halden.example:
ignore_rules: [aws_access_key_id]
# Tools that pin their own certificates: never terminated, reported as exempt.
exempt: [api.gofile.io]
websocket: relay
large_bodies: stream
record_path: findings_only6. Tell the people covered
The first-run sheet asks whether you've told the people whose agents this covers; it is not required. The works-council pack states what inspection reads, keeps, sends and for how long: requests' paths, queries and bodies are read in memory on the machine, never responses or headers; a finding keeps the site, the rule, a count, where and what was done; the redacted path stays on the machine and never leaves it, not even as a hash; Central Lite receives findings, uninspected records, alerts and ten-minute counts per site, with the site as a keyed hash; findings are kept as security records and counts as operational ones.
Under the non-monitoring profile no path is kept at all, on the machine included; findings and alerts reach Central Lite without the site, and the counts per listed entry as you wrote it (*.hooks.halden.example), with no task or person.
agentgarm compliance pack --out works-council-packSee also
Prepare a works-council or DPIA packGenerate templates that describe what your deployment collects, keeps and sends, and switch the fleet to the non-monitoring profile during talks.