ProPR
Log in Get ProPR

Put recurring agent work on a schedule.

An automation is a saved prompt that ProPR runs on demand or on a schedule. Each run is an isolated ProPR task run that researches your repositories or the web and writes a Markdown report. You decide whether the report stays a report or becomes work.

Anatomy of a run

  1. Trigger. Run now, a cron schedule, an HTTP call, MCP, or propr automation run.
  2. Report. The agent researches in an isolated run and writes Markdown. It never commits or pushes.
  3. Act, if you allow it. A second run turns the report into tasks, plans, goals, TODOs and comments through ProPR's MCP tools.

Product path overview

A saved prompt with a job to do.

Write the instructions once. ProPR runs them with the same isolation, models and receipts as every other task.

Four automations with their coding agent, schedule and last run: one awaiting approval, one deferred
Each automation shows its agent, schedule and how the last run ended.
The weekly dependency review reads two repositories, northwind/storefront-web and northwind/orders-api; up to ten are allowed

The definition

A name, a prompt, the coding agent and model, and up to ten repositories. An automation without a repository works from its prompt, input files and the web.

The automation prompt names dependency-policy.md, attached below it as an input file

Input files

Attach up to ten files: a checklist, a dependency policy, a list of competitors. Each run finds them in its workspace, listed in its prompt.

Previous reports switched on: the last two reports feed back into the next run

Memory across runs

Feed the last one to five reports back into the next run, so a weekly review compares this week with the last.

Five ways to start a run.

One trigger primitive serves all of them, so a scheduled run and a run from chat produce the same record.

  • Run now

    Start a run from the web UI. If the agent's subscription is close to its limit, ProPR shows the usage and lets you run anyway.

  • HTTP

    POST /api/automation-definitions/:id/runs starts a run from GitHub Actions, a webhook relay or an external cron. Send an Idempotency-Key header and a retried call returns the same run.

  • MCP

    Ask your chat client to run an automation, read its report, and approve or reject the next step. MCP from chat

  • CLI

    propr automation lists automations, starts runs, waits for them, prints reports to stdout, and approves, rejects or cancels runs.
Run history for one automation: runs started by schedule, MCP, the CLI from GitHub Actions, the HTTP API and Run now
Schedule, MCP, CLI, HTTP and Run now land in the same run history.

Schedule

Five-field cron or @hourly, @daily, @weekly, @monthly, evaluated in UTC, at most every 15 minutes. Presets cover hourly, daily at 09:00, weekdays at 09:00 and Mondays at 09:00, and the editor previews the next run. Missed slots collapse into one run.

A UTC cron schedule with presets for hourly, daily, weekdays and Mondays, and a preview of the next run
Cron presets, evaluated in UTC, with the next run previewed.

From a report to actions, at the autonomy you choose.

ProPR never parses the report for commands. An optional acting step reads it in a second isolated run and acts through ProPR's MCP tools.

Autonomy set to Preview & approve: the acting step waits for approval before it uses ProPR tools
One switch per automation decides what happens after the report.
Monday 09:00: report, approval, tasks.A weekly dependency review fires on its schedule and reports in Markdown. You approve with a note, and the acting step files two tasks and a TODO through ProPR's MCP tools. The tasks start right away.
A completed dry run of the competitor scan, started from chat: the report is saved and the run ends there

Dry run

The default. ProPR saves the report for you to read, and the run ends there.

A preview-mode run awaiting approval, with the note "Only fix auth-tokens." for the acting agent

Preview + approve

The acting step waits for you. Read the report, then approve it with an optional note for the acting agent, or reject it and keep the report. Automations with a repository also notify your Inbox.

An Auto run of the issue triage: the acting step filed a task, commented on two duplicates and added a TODO, with no approval step

Auto

The acting step runs as soon as the report is ready and creates tasks, plans, goals, TODOs and pull-request comments on its own.

Hard limits on the acting step.

The acting step holds a short-lived ProPR MCP grant with read, plan and execute scopes, limited to the automation's repositories. ProPR revokes it when the step ends. The grant appears in Connected apps, and the MCP Log records every call.

  • Acts through ProPR only

    It creates tasks, plans, goals, TODOs and pull-request comments, and checks existing tasks and TODOs first to avoid duplicates. Without a repository, it creates TODOs only.

  • Cannot ship anything

    It cannot merge pull requests, deploy, change settings, or edit, commit and push code. New code still arrives as a task, a pull request and your review.

  • Cannot run itself

    It cannot trigger, approve or reject automation runs, so one automation never sets off another.

  • Treats the report as data

    The acting agent reads the report as input to evaluate. Instructions inside it carry no authority.
What the acting agent did after approval: one task for auth-tokens, exactly as the note asked, after checking for duplicates
The acting step filed one task, as the note asked, after checking for duplicates.

Give each automation only what it needs.

Every automation reads its repositories. Add web access or ProPR tools only where the job calls for them.

Per-automation capabilities: read repositories and web access on, ProPR tools off
Repository reads on by default; web access and ProPR tools switched per automation.
CapabilityWhat the report step getsCoding agents
Read repositories (default)A read-only checkout of the selected repositories. Without it, the agent starts in an empty workspace with no source code.All five agents
Web accessWeb search and page fetches while it researches.Blocked natively when off on Claude Code and Codex. Antigravity CLI, OpenCode and Mistral Vibe receive the restriction as an instruction.
ProPR toolsRead-only ProPR MCP access to tasks, goals and plans.Claude Code and Codex

Preview + approve and Auto need an agent that uses ProPR's MCP tools: Claude Code or Codex. Other agents run automations as dry runs. ProPR tools and the acting step also need the instance's MCP server turned on. To give an automation your own MCP servers and skills, select an extension profile.

Unattended runs respect your subscription.

Before a scheduled, HTTP, MCP or CLI run starts, ProPR checks the agent's capacity in Agent Tank against a pause threshold, 90% by default.

  • Weekly allowance past the threshold

    ProPR skips the run.

  • Session window past the threshold

    ProPR defers the run until shortly after the window resets, and skips it after six deferrals.

  • Auto mode holds back

    If the gate stops an automatic acting step, the run waits for your approval with the reason attached.

  • Monitoring never blocks work

    With Agent Tank disabled, unreachable or out of date, runs proceed. Run now always proceeds after its warning.
A scheduled run deferred because the agent's session usage is above the 90% pause threshold
A scheduled run deferred until the session window resets.

Four automations to start with.

The weekly dependency review in preview mode: storefront-web and orders-api on Claude Opus 5.5, every Monday at 09:00 UTC, its last run waiting for review

Weekly dependency review

Every Monday the agent reviews the repository's dependencies and reports what needs attention. In preview mode you approve with a note such as "Only fix auth-tokens", and the acting step files exactly that work.

A daily digest of blocked pull requests: what each one waits on and for how long, and what changed since yesterday

Blocked pull requests

"Review open pull requests in these repositories and summarize which ones are blocked and why." A digest of stalled work that compares each run with the previous report.

An issue triage report: one payments issue to act on first, two likely duplicates, two issues missing reproduction steps

Nightly issue triage

Summarize new issues overnight, then ask your chat client what last night's triage run reported. A GitHub Actions step can trigger the same automation and save its report to a file.

The competitor scan's capabilities: web access on, no repository to read

Competitor scan from chat

An automation with web access and no repository. Tell your chat client to run your competitor scan now, and read the report when it lands.

Trigger it from CI.

The CLI waits for the run and prints the Markdown report, so a workflow step can save it. The exit code says whether the run completed, timed out, was skipped, or is waiting for approval.

npx propr-cli automation run "$AUTOMATION_ID" \
  --idempotency-key "$GITHUB_RUN_ID" \
  --source github-actions \
  --wait > report.md

Create in the web UI.

You create and edit automations in the web UI or over the REST API. MCP and the CLI start runs, read reports and decide on the acting step. Each automation belongs to one owner, and schedules run in UTC.