Start it any way
Run now from the web UI, a UTC cron schedule with presets, an HTTP call with an Idempotency-Key, your chat client over MCP, or propr automation run from CI.
ProPR 0.9.0 · October 2026
Goals, Automations, and restricted agent networking.
ProPR 0.9.0 is the largest release since the first public version. Agents take on standing objectives, recurring work runs on a schedule, agent containers get a tighter network and credential boundary, and ProPR reaches you on the desktop, in the browser's notification tray, and in your chat client.
Overview
New
Goals: give an agent an objective, then steer it.
A task answers one request. A goal keeps a coding-agent session working toward an objective until it is done, and you stay in the loop the whole way.
Two ways to run
The agent implements directly while ProPR owns every commit and checkpoint on a draft PR, or the agent orchestrates: it breaks the objective into issues and drives them through ProPR's normal task path.Corrections mid-run
Send an instruction or a question at any time. The goal timeline records it verbatim. Pause, resume, cancel, or change the model without losing work.Native where the agent supports it
Claude Code and Codex run goals natively with live steering. Antigravity CLI resumes its own conversation and takes corrections at safe boundaries.
New
Automations: recurring agent work, on a schedule.
An automation is a saved prompt that ProPR runs as an isolated task. Each run writes a Markdown report, and you decide whether that report turns into work.
A report with memory
The agent reads the selected repositories, the web if you allow it, its input files and up to five previous reports, then writes Markdown. Report runs never commit or push.
Actions at your autonomy
Dry run is the default. Preview + approve waits for your approval; Auto acts at once. The acting step works only through ProPR's MCP tools and cannot merge, deploy, change settings, or trigger or approve runs.
How Automations work · Automations guide
New
New Task: one instruction, straight to a PR.
Small, clear changes skip the plan. New Task sends one instruction straight to implementation with the repository's settings.
Still traceable
ProPR files a GitHub issue and runs the normal implementation task, so the issue, task and pull request link up exactly as planned work does.The create button follows you
The primary action is New Task by default, New Plan in Plans and Planner Studio, and New Goal in Goals. The other two sit in its menu.Start from anywhere
A repository's New task prefills the repository, and a to-do's Run task prefills its text. New Task remembers the repository, agent and model from your last submission. Attached images render inline in the GitHub issue.
New
Safer execution: restricted network and read-only tokens.
Agent containers get a tighter boundary on what they reach and what they change. ProPR creates every branch, commit, push and pull request itself.
Restricted network mode
Optional, and off by default. Restricted mode starts the agent container with no network. Its only route out is a per-run proxy on the worker that allows the agent's model provider, GitHub, npm and PyPI, plus the hosts you add. It needs no elevated container privileges.
Instance and repository policy
Set the mode and allowed hosts in Settings, and enforce restricted mode across the instance if you want. Repositories choose their mode and extra hosts under network: in .propr/workflow.yml; enforcement overrides a repository's open.
Denied hosts on the timeline
Each restricted run lists every denied host, with attempt counts, on the task timeline. A missing allowlist entry is one look away.
Where it applies
Implementations, follow-ups, /fix, reviews, goals and indexing runs. Plan generation runs on the open network. Restricted mode needs a Linux Docker host, and Antigravity CLI falls back to open mode, or is refused when restricted mode is enforced.
Fails closed
If ProPR cannot read the network policy, the run fails. If the worker stops, its restricted containers lose network. The allowlist matches hostnames, so an agent still reaches any allowed host it holds credentials for.
Read-only agent tokens
Implementation, follow-up, review-fix and direct-goal agents get a read-only GitHub token, and the repository's Git data mounts read-only. Orchestrating goals keep write access to drive their issues.
Execution safety documentation
New
Repository rules live in the repository.
Two optional files in .propr/ let a repository own its workflow. ProPR reads them from the base branch, so every change to them goes through review.
Lifecycle hooks
.propr/workflow.yml runs your commands around the agent: after_create, before_run, after_run and before_remove. Add an instructions file, validation commands that ProPR runs and reports after the agent exits, preview types, and a repository-wide limit on parallel tasks.
Network, spend and merge
The same file sets the repository's network mode and allowed hosts, a per-run spend cap, and its auto-merge policy. An invalid file fails the run before the agent starts.
Pull request templates
.propr/pr-template.md shapes the title and body of the PRs ProPR opens, with placeholders such as {{issue_number}}, {{model}} and {{cost}}. Without one, ProPR appends the repository's GitHub PR template.
propr init scaffolds both files with comments, and propr repo validate checks the template locally. Repository workflow · PR templates
New
Guardrails for runs nobody is watching.
Four controls bound what an unattended run costs, how long it hangs, and what merges without a person.
Protected paths for auto-merge
The auto_merge block in .propr/workflow.yml decides whether ProPR arms auto-merge and with which method. A PR that touches a protected path, or anything under .propr/, waits for a human merge. When ProPR cannot read the policy or the diff, it skips auto-merge and says so on the PR.
Spend caps
Cap a run's estimated cost per task, per repository or for the whole instance. At the cap ProPR stops the run, publishes its partial work, and notifies you in the Inbox and on the PR.
Stall watchdog
ProPR stops an agent that goes silent for 10 minutes, 30 during a running tool call, or streams empty output, and publishes its partial work. Before, a hung agent held its worker slot until the 24-hour timeout.
Replacement runs
When a worker dies mid-task, ProPR starts one replacement with the same agent and model, continuing from the pushed branch. A transient provider error gets up to two replacements by default. Rate limits and credential errors never trigger one.
Redesigned
Tasks becomes a triage console.
One row per task, with status, agent, duration and review score. A run-count chip sums up earlier runs, so a long PR history takes one line.
Open it beside the list
On screens 1280px and wider, a task opens in a pane next to the list and its link goes in the URL. j and k move between rows, and Esc closes the pane.Every run on one timeline
Implementation, reviews, fixes and Ultrafix cycles list in order. Pick a run to switch the screenshots, trace and changed files to it.Scores over time
ProPR keeps every review score, and the task shows the PR's score trend.Commands on the newest run
/fix,/reviewand/ultrafixappear only on a task's newest finished run, so an old command can't be replayed.Find anything
Cmd/Ctrl+K opens a command palette with a live preview of tasks, plans and repositories.
New
Task assignment, kept in sync with GitHub.
Every task is assigned through the issue or pull request behind it, so ProPR and GitHub always agree on who owns the work.
Filter by owner
The Tasks screen shows assignees and filters by Assigned to me, Unassigned or chosen people. propr task list --assignee me and MCP list_tasks take the same filter.
Hand off finished PRs
Per repository, ProPR assigns each pull request when its work completes, to a default assignee or the source issue's author, and can request their review. It never removes an existing assignee.
The assignment gate
With the gate on, an assigned pull request starts follow-up work only from its assignees' comments. Unassigned PRs, failed-CI follow-ups and the Ultrafix loop are unaffected.
Goals, plans, automations and to-dos also show who created them. How assignment works
New
Agent extensions: your MCP servers and skills in task runs.
An extension profile gives Claude Code and Codex task runs a chosen set of MCP servers and skills. Select one on New Task or an automation.
Listed tools only
Add a remote MCP server, discover its tools, and allow the ones the agent may call. Write servers need the run's explicit write authorization.
Credentials stay out
Bearer tokens stay encrypted in ProPR. The agent reaches each server through a per-run gateway that closes when the run ends.
Pinned skills
Import skills from a GitHub repository or an archive. ProPR pins the commit and digest and mounts them read-only. Every profile edit is a new revision.
New
ProPR Desktop for Linux and macOS.
The desktop app runs the ProPR web UI in a sandboxed window and connects to one or more ProPR instances.
Browser-approved pairing
Enter an instance address, approve in your browser, and confirm the GitHub account in a native dialog. Tokens live in the OS credential store.Saved instances and accounts
Keep several instances and several GitHub accounts, even two accounts on the same instance, and switch between them.Native menus and tray
A keyboard shortcut for every menu action, an application menu button on Linux, a tray with task and plan counts, and native notifications you switch on per device.Guided local setup on Linux
Set up this computer signs you in to GitHub through your browser, starts a local ProPR stack with the same setup engine as the CLI, then opens browser pairing.
New
Inbox, installable app, and push notifications.
ProPR events land in one Inbox, and the ones you choose reach your phone or desktop as push notifications.
One Inbox
Plan, task, review, pull request, indexing and system events, each opening the task, plan or goal it belongs to. Read and dismiss state syncs across tabs. Dismissing never cancels work.Triage at a glance
On wide screens the Inbox is a queue with a verdict beside it: the review score and findings, a short diff, and the next command. j/k step through, x dismisses and opens the next item.Install it as an app
Install the web UI from your browser and enable push on each device. On iPhone and iPad, add it to the Home Screen from Safari first.Your rules
Choose Inbox and Push per category, set quiet hours, and mute a noisy repository. Muting stops notifications; automation keeps running.
/fix and x clear an item and open the next; j and k step through.
New
Run ProPR from your chat client over MCP.
ProPR 0.9.0 ships an authenticated MCP server at https://<your-instance>/api/mcp. Turn it on in Settings and set the scope ceiling.
You grant, you revoke
Each connection goes through a consent screen with a repository picker, and optional scopes start unchecked. Connected apps lists every grant with its repositories and last use, and revokes it in one click.
The whole workflow
Ask what is running and what needs you. Create and refine plans, launch goals and tasks, run automations and approve their next step, review pull requests with the models you pick, fix selected findings, and resolve merge conflicts. Search and read code in granted repositories without cloning.
Receipts you can follow
Long operations return receipts that move from accepted to running to a final state. A work overview joins tasks with their PRs, reviews and checks in one call.
Answers from the shipped docs
Docs tools search the documentation bundled with your release, so answers match the version you run. A settings finder tells you where each setting lives.
An access log for admins
The MCP Log records tool names, outcomes and timing. Arguments and result bodies stay out of it.
Your infrastructure
Agents, clones and task state stay on your server. Requested results travel to the chat client and its model service, as any chat answer does.
New
Plan history and safer refinement.
A plan is a working document, and now it keeps its past. ProPR stores the plan before every generation, refinement and edit.
Restore any version
Open Plan history, preview a saved version, and restore it. Restoring saves the current plan first, so nothing is lost.Refinement returns the whole plan
Ask for one change and get the complete plan back, unchanged tasks included.No silent partial plans
Generation validates every issue before ProPR accepts it. Incomplete output fails with an error and leaves your saved plan alone.
New
Visual previews on pull requests.
Review UI changes by looking at them. The agent captures the changed screen, and ProPR posts it on the PR. How visual previews work
Only when it is visible
The agent captures images or video only when the change shows up on screen, using the repository's own tools: Playwright, Storybook, a browser or an emulator. Follow-up commits get fresh media.
Inline in GitHub
Previews publish as GitHub attachments, so they render in the pull request. Task and goal pages show them in a zoomable gallery.
Full-resolution originals
With Connect Plus, ProPR also keeps each full-resolution original in managed storage and links it from the PR.
New
Synthetic pools: one model name, many accounts.
A synthetic pool is a virtual agent and model backed by several direct agent accounts. Repositories, the planner and reviews point at the pool, and ProPR decides which member runs each task.
Priority tiers
Put the primary account at the top and a fallback below it. Within a tier, route round robin or by remaining usage.Usage caps
Cap a member at a share of its session or weekly allowance, measured by Agent Tank.Failover
A retryable provider error moves the same task to the next member in the same workspace and branch.Honest attribution
The UI shows the pool; task details and the LLM log show the account and model that actually ran. Manage pools in Coding Agents or withpropr agent pool.
Improved
Agent Tank now ships inside ProPR.
Agent Tank shows how much subscription capacity each agent has left. In 0.9.0 it is bundled in the ProPR agent image: pick Bundled in Settings and ProPR runs it on demand with the credentials your tasks already use.
Three modes
Disabled (the default), Bundled (recommended), or External for an Agent Tank daemon you already run. Upgrades keep an existing daemon connection.Covered agents
Claude Code, Codex and Antigravity CLI. Their usage bars show reset countdowns, and each LLM log entry records the usage it consumed.Capacity drives automation
Scheduled automations skip or defer when an agent nears its limit, and Ultrafix escalation passes over models that are out of capacity.Usage history
Hover a usage row in the sidebar to see that limit over its last two periods. Analytics charts every provider's usage from 24 hours to 14 days.One command
propr tank bundledswitches it on from the CLI.
Improved
Sharper review commands and CI control.
Pick exactly what to fix
Reviews number merge blockers F1, F2 and optional suggestions S1, S2 across the whole PR. /fix applies the pending blockers, /fix F20 S3 S5 applies exactly those records, and /fix all applies every pending finding and suggestion. Each review ends with a copyable /fix line to trim. One malformed ID rejects the whole command, so nothing half-applies.
Cancel superseded CI
Pick the workflows to cancel for each repository. When a follow-up starts implementing, ProPR cancels those runs on the exact PR head being replaced, and restarts them if no new commit lands. Closing a PR uses the same policy.
Checks that never block
Mark check-name patterns as non-blocking. Their failures stop holding up /merge, auto-merge and Ultrafix, and GitHub still shows the result.
Epic progress on the PR
As child PRs merge, the epic PR keeps one comment counting how many have landed, and posts a confirmation when the last one merges.
Epics that keep moving
When a repository disallows GitHub auto-merge or a PR has no CI, ProPR merges the plan's PRs itself once every gate passes, and suggests the repository setting.
One CI-wait comment
Ultrafix's "waiting for CI" comment updates in place: it becomes the next review, or the stop notice if CI never settles.
Updated
New default models and refreshed agents.
Claude Code
Claude Opus 5.5 is the new default. Claude Fable 5.1, Sonnet 5.5 and Haiku 5.5 join the catalog, and the plain fable, sonnet and haiku aliases select them. Older models move behind the legacy fold.
Codex
GPT-6 Astra is the new Codex default. GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna join the catalog.
Antigravity CLI
One entry per model with a separate reasoning selector: Gemini 3.8 Flash (the default), Gemini 4 Argon (recommended for planning and review), Claude Sonnet 5.5, Claude Opus 5.5 and GPT-OSS 120B. Gemini 3.1 Pro is retired, and its label runs on Argon. Goals run through resumable conversations.
Mistral Vibe
GLM 5.3 and GLM 5.2 join Mistral Medium 3.5 and run with the same Mistral credentials. Devstral configurations move to Mistral Medium.
OpenCode
Big Pickle becomes the default, and the free lineup adds Muse Spark 1.2 and 1.3, Nemotron 3.5 Lightning, and Ling 3.0 Flash Fin.
Pinned CLIs
Every agent image ships an updated, pinned CLI: Claude Code 2.1.293, Codex 0.160.0, Antigravity CLI 1.4.0, OpenCode 1.18.31 and Mistral Vibe 2.25.8.
Redesigned
A dashboard that starts with what needs you.
The dashboard now opens on Needs attention: failures and decisions waiting on a human. Happening now shows running tasks and goals with their current step and last output. Completed groups finished work, and earlier updates load on demand.
Live, push-driven
Updates arrive over the socket, and hidden tabs reconcile when you return.Analytics
Analytics reports runs per task, first-time pass rate, time to merge, autonomy, prompt-cache hit rate, and agent efficacy by model, over any window from 24 hours to all time. Historical stats on the dashboard read the same figures.Access
Administrators assign member and admin roles, with an audit trail, under Access. Adding or removing someone offers to update the GitHub trigger whitelist to match.
Upgrading to 0.9.0.
Back up the database and .env, then update the CLI and restart the stack. The CLI pins every service and agent image to its release.
npm install -g [email protected]
propr start --restartLauncher installs pull the new propr/launcher image and restart. The maintenance guide covers backups, pinning and rollback.
What to switch on
Off until you enable them
MCP, visual previews, CI cancellation, restricted network mode and Agent Tank start disabled. Existing Agent Tank daemon connections carry over as External. Automations that act through ProPR tools need MCP enabled.Existing GitHub Apps
Subscribe your App to the Check suite event so Ultrafix recovers from red CI.propr check --verifyreports it when it is missing.Per person, per device
Push notifications, desktop notifications and the experimental voice briefings are opt-in in each user's Settings.Desktop is separate
ProPR Desktop installs on your workstation and connects to any 0.9.0 instance. Updating it does not upgrade the server.