A lawyer sorted his own skills in one sentence: anything with a yes/no gate needs a person, anything that only reads can run alone. The three-question test, his skills sorted, and what the server actually changes.
[ KEY TAKEAWAYS ]
A Claude skill can run on a schedule when three things are true: everything it reads is reachable from the cloud (a connector on a server, not on a laptop or a synced folder), it only reads or makes writes you would accept without looking first, and no step needs a human to answer a question. A weekly timekeeping report and a sweep for matters nobody has touched pass all three and can run at 7am on Monday with every laptop closed. A Complaint drafter that reads the file and then asks you to confirm the defendants and the court fails the third on purpose, and no server changes that. Cowork scheduled tasks run on Anthropic's side, so they reach a remote connector and never a desktop one.
questions that decide it for any skill, on any assistant with scheduled runs.
of his skills run alone: the weekly timekeeping report and the quiet-matters sweep.
stays gated by design: the Complaint drafter. The gate is the feature.
Monday, laptops closed, is when a scheduled skill runs once the connector is on a server.
A lawyer with a dozen Claude skills asked us this week whether moving the connector off everyone's laptop would help him. He answered his own question before we could, and his answer is the clearest version of this we have heard, so here it is with the details filed off.
"It should go look at all the underlying documents, then the notes and custom fields, then ask the user a series of questions. It confirms who the defendants are going to be. It confirms what court you are filing in. It requires a gate where the user has to say yes. This is to prevent AI slop." A lawyer describing his Complaint drafter, and then concluding, correctly, that hosting it would not help.
Ask three things of any skill. First, can everything it reads be reached from the cloud? A case system through a connector on a server, a document system through its vendor's cloud connector, a payroll system through its connector: yes. A connector installed on a laptop, a folder synced to a desktop, a file on a network drive: no. Scheduled tasks run on the vendor's servers, which can reach a URL and cannot reach a closed laptop. This is the part most firms miss, and the reason the connector on your laptop is invisible to Cowork.
Second, does it only read, or does it make writes you would accept without looking? Pulling time entries is a read. Posting a score into a custom field you review on Monday is a write most firms accept. Creating a matter, assigning tasks to people, moving a case into a new stage: writes you want to see before they happen.
Third, does any step need a human answer? "Which court?" "Is this the right client?" "Is this the current version?" If the skill stops to ask, it cannot run at 7am, and it should not.
Needs a person
Complaint drafter
Reads the documents in the document system, then the notes and custom fields in the case system, then asks a series of questions. Confirms the defendants. Confirms the court. Only then drafts, then saves. Fails question three on purpose. A server changes nothing about it except that each person runs it from their own login and the run shows in one log.
Needs a person, mostly
Intake scoring
Reads the intake notes and the vetting fields, scores the case, lists red flags and missing documents, writes the score back. The reading and scoring could run alone. Whether the write-back is accepted unattended is a firm decision. Most keep a person on it until the scoring has been right for a month.
Runs alone
Weekly timekeeping report
Pulls everyone's time for the week, checks for gaps, entry quality and billable share, and produces a dashboard plus a card per person. Reads only. Nothing to confirm. He runs it by hand today because the connector is on a laptop; on a server it is a Monday 7am task. The recipe is here.
Runs alone
Quiet-matters sweep
Every open matter with no note, no time entry and no calendar entry in sixty days, quietest first, then a look in the document system to explain each one. Reads only. He checked it by hand against a hundred live matters before trusting it, which is the right order.
It is tempting to read "needs a person" as a problem to engineer away. It isn't. His words: once the thing gets off in the wrong direction, it starts producing material that has no relationship to reality. The gate exists so that the defendants and the court are confirmed by someone who will sign the filing, and so that a wrong assumption dies at step two instead of appearing in a draft at step nine. Automating past that gate would make the skill faster and the output worse.
So the honest split is this. Hosting helps the skills that were already safe to run alone. It does not help the skills that need judgement, and nobody should sell you a server on the promise that it will.
Writes change the risk class A skill that only reads can be wrong and cost you nothing but a minute. A skill that writes without a gate can be wrong in your case system. Claude's scheduled-task setup has an approval mode, and most custom connectors gate every write unless an admin allows it for all tasks. Decide on purpose which writes you accept unattended, and keep the list short.
Three things. The connector to your case system has to be reachable over the network: a server in your firm's cloud or a hosted one, added to Claude as a custom connector, with each person signing in as themselves. The skill has to be saved where the scheduled task can find it, which in Cowork means your Claude account rather than a folder on a laptop. And the write permissions have to be decided in advance, because the task form has an approval mode and a custom connector's write tools are gated behind approval unless an admin allows them for all tasks. A read-only skill sails through. A skill with one write in it stalls at the approval prompt at 7am and nobody is there to click.
Our open-source connector marks its read tools as read-only in the way Claude expects, so a report skill does not trip that gate; others may not, which is worth checking with one test run before you trust the Monday slot.
Which of your skills would run alone?
Send us the list of what your firm already does with Claude and we will sort it with you on a 30-minute call: what can be scheduled this week, what stays gated, and what the connector needs to be for either.
Gated skills (drafting, filing, anything that asks)
Same skill, same questions, same person. Gains: each person runs it from their own login, the run appears in one firm-wide log, and nobody installs or updates anything on a laptop. No automation gain, and that is fine.
Read-only report skills (timekeeping, quiet matters, deadlines)
Become scheduled tasks. The report is waiting when the first person arrives. Nobody remembers to run it, so it always runs.
Read-then-write skills (scoring, tagging, routing)
Schedule the read, decide the write. Either accept that specific write unattended and allow it, or leave the write for a person and schedule only the part before it.
If you go through your own list with the three questions, you will probably find the same two-to-one split he did. Most of what a firm does with AI needs someone in the chair, and the few things that don't are exactly the things nobody has time to run by hand.
Yes. Cowork scheduled tasks are available on all paid plans and run hourly, daily, on weekdays, weekly or manually. Anthropic's help centre says they run remotely, on their cadence, even when your computer is asleep or the desktop app is closed, and that they use the connectors and plugins you have set up. A task that needs local files or apps runs only locally.
Because the task runs on Anthropic's servers and the connector is a process on your laptop. A server cannot reach into a closed laptop. For a skill that reads your case system to run on a schedule, the connector has to be reachable over the network: on a server in your firm's cloud or hosted by a vendor, added to Claude as a custom connector.
Any skill where a wrong answer at one step changes everything after it: naming the defendants, picking the court, deciding which version of a document is current, choosing what to file. Those steps exist to stop the AI drifting from reality, and a person has to answer them. Put the skill behind a gate and run it when someone is there.
Split it in your head. The reading part can be scheduled. The writing part depends on what it writes. Posting a score into a custom field you review later is a write most firms accept unattended. Creating a matter, assigning tasks to people or filing anything is not. Claude's scheduled-task setup has an approval mode, and most custom connectors gate every write behind approval unless an admin allows them for all tasks, so an unattended run with writes in it will usually stall rather than act.
Barely. The Complaint drafter runs the same way, with the same questions, from whoever's login runs it. What changes is that each person signs in as themselves, the run shows in one firm-wide log, and nobody has to install or update anything on a laptop. The automation gain is on the ungated skills only.
Yes. Any assistant with scheduled runs has the same shape: the schedule runs on the vendor's side, so it can only reach systems exposed over the network, and anything that needs a human answer mid-run has to wait for a human. The three questions do not depend on the vendor.
NEWSLETTER
[ WRITTEN BY ]
Petar Jovanović
Co-Founder & Technical Lead
Co-Founder and Technical Lead at Oktopeak. Builds regulated software for legal and healthcare teams, and leads the rescues of codebases other vendors left half-finished.
[ LEGAL TECH ]
LEGAL TECH
Aug 27, 2026 · 11 min read
Claude Cowork scheduled tasks run on Anthropic's servers, so a Clio MCP server installed through claude_desktop_config.json is invisible to them. Five ways to run a Claude skill on a schedule, sorted by where the connector has to live, with the trust requirements of each.
Read ArticleLEGAL TECH
Oct 9, 2026 · 9 min read
Every Monday at 7am: hours per workday per person, gap days, billable share, entry quality, unbilled time, with time off from the payroll system so a day off isn't flagged as missing time. Reads only, nothing written back.
Read ArticleLEGAL TECH
Oct 9, 2026 · 8 min read
A managing partner can't tell who uses Claude on Clio from five laptops. What each setup records, where the one firm-wide log lives, what Claude's own admin tools show by plan, and five questions for any vendor.
Read Article[ GET STARTED ]
30-minute call. No pitch deck. Just an honest conversation about your project.
