Agentic Development Environment
"I wanted my agent to work all night, but it stopped 2 minutes after I went to sleep to ask a stupid question!"
Use SLAM.
"I wanted my agent to get some work done while I ran an errand, but it got blocked on some stupid environment issue and stopped working."
Use SLAM. It self-corrects.
"After a couple weeks working on this project, my agents waste all their time on useless tests, documentation, comments and bad architecture. Development is slowing to a crawl and my subscription quota is wasted."
Use SLAM. And tell your XO to cut the crap.
"I'm sick of having these long planning conversations with AIs, ending up with disorganized implementation plans and wearing me out."
Use SLAM to create a slate of nicely organized tickets, then run SLAM Auto Mode to execute them quickly.
"I'm working 12 hour days just pressing Enter."
Use SLAM.
"My boss is pressuring me to ship faster and faster."
Use SLAM.
"I'm on a $100/mo subscription plan but I can't use all of it. I'm leaving money on the table."
Use SLAM with Auto Mode. It'll burn your entire subscription, I promise.
"I want to manage agents from multiple providers on multiple projects simultaneously while out and about."
SLAM Remote is a planned future subscription at a proposed $5/mo, not available yet. (Local SLAM is free.)
Product workflow
Tickets record the work to do. Planning captures decisions before an agent implements a ticket or phase. SLAM handles the ticket status and commit bookkeeping.
Review the plan and resulting changes. Implement runs one phase or one unphased ticket and ends at done for review. Auto Mode launches a fresh agent for each phase, commits completed work, and selects the next ticket.
The Plan Tab organizes tickets and launches work. Explore browses files and symbols, Track reads diagnostics, and Review walks diff hunks. Research-Plan-Implement is an existing workflow that inspired SLAM.
See the interface: Explore Tab.
Illustrative example / no agent is running
Example ticket: preserve a player's sound setting. Every artifact below is fictional. Advance or revisit the steps to see how a ticket keeps findings, human decisions and verification together beyond a conversation.
Example ticket / status: open
Request: Remember my sound setting after restart.
Finding: Settings.load() always uses sound = true.
Existing storage: settings.jsonThe finding stays beside the request so the next session can start from evidence rather than reconstructing the conversation.
Question: What should happen on a first launch?
Human decision: Sound on. Preserve an explicit off.
Plan:
- Save sound in the existing settings file.
- Default to true only when the value is missing.
- Check missing, true and false values.
- Leave volume controls outside this ticket.
Status after planning: plannedThe human settles the default and scope before implementation. The decision and acceptance checks remain in the ticket.
See the interface: Plan Tab / Ticket detail.
One unphased example ticket:
planned -> started -> done
Illustrative change:
- sound: true
+ sound: saved.sound ?? true
Illustrative verification record:
missing -> on; true -> on; false -> offAn Implement call ends at done. In Auto Mode, the supervisor launches a fresh agent for each phase, commits completed work and picks the next ticket. This example only displays artifacts; it executes no code or tests.
See the interface: Auto Mode.
Demonstration of hunk inspection in the Review Tab. These fictional changes implement the recorded human decision and acceptance checks: sound on when missing, preserve explicit off, and leave volume controls outside scope.
@@ -1,2 +1,2 @@ load(): Settings
Old line 1, new line 1, context: const saved = readSettings("settings.json");
Old line 2, removed: return { sound: true };
new line 2, added: return { sound: saved.sound ?? true };The nullish default preserves false. Using saved.sound || true would turn an explicit off back on. Acceptance checks: missing → on; true → on; false → off.
@@ -5 +5,3 @@ save(settings: Settings): void
Old line 5, removed: writeSettings("settings.json", {});
new line 5, added: writeSettings("settings.json", {
new line 6, added: sound: settings.sound
new line 7, added: });Save the current boolean in the existing file so an explicit off survives restart. Human inspection checklist: switch sound off, save, restart, and confirm it stays off. Volume controls remain outside this ticket.
Notes live only in this page until reset or reload. This example executes no code and submits no review.
Review walks diff hunks. Compare both changes with the ticket before deciding whether to accept them. Auto Mode committing work does not replace human judgment. This teaching example does not demonstrate queued edits, discards or a review agent.
Fictional illustration — no agent is running. Advance manually to follow one successful unit of work.
Auto Mode implements one phase or an unphased ticket, commits the completed unit, then selects ready work from its queue. Rank expresses order; readiness also depends on unfinished dependencies. A human hold prevents the next launch.
Completed earlier: #A100 — Default sound on · Status: committed.
Hold example: the human places the supervisor on hold before another launch. The illustration pauses at its current step; release the hold to continue, or reset to the initial queue. Hold is a supervisor state, not a ticket status.
Product gallery
Actual SLAM interface, captured in a seeded demonstration workspace. These stills show demonstration data, not customer activity. No agent is connected.
Open a full-size image to read the terminal text. Scroll within the viewer to inspect it; Close or Escape returns to the gallery.
Back to the explanation: Plan Tab.
Back to the explanation: Ticket detail.
Back to the explanation: Auto Mode.
Back to the explanation: Explore Tab.
Installation
Install from a checkout with Node.js 20 or newer, git, and a working C toolchain. npm install builds tree-sitter's native bindings. SLAM launches an agent CLI you already have; install and configure your agent client separately.
git clone https://github.com/LonePineGames/SLAM.git slam
cd slam
npm install
npm linkKeep the checkout in place: the linked slam command runs TypeScript from it. There is no published npm package to install.
cd ~/work/my-project
slam init
slamRun these commands from your own project directory. slam init fills gaps in git setup, tickets, agent guidance and the project's .mcp.json entry, then exits; slam opens the TUI on the Plan tab. The first launch after a fresh clone may pause once to build rtui.
The commands above remain selectable if copying is unavailable. Replace ~/work/my-project with your repository path before running the second block.
Remote
SLAM Remote is a planned future subscription at a proposed $5/mo, not available yet. (Local SLAM is free.)
SLAM Remote connects a browser to SLAM's terminal grid. Run slam remote to manage paired devices and the repositories this machine allows remote clients to open. A browser can mirror a terminal or start a session in an allowed repository.
Back to homeInteractive demonstration. All host, device and repository labels are fictional. These controls only change this illustration: no relay, login, pairing, agent execution or network requests.
Run slam remote on the host to inspect paired devices. Renaming a device changes its label; revoking one removes its access.
This represents devices already paired from elsewhere. A browser on the host itself does not need pairing.
On the host, r in slam remote opens the repository list. The host explicitly adds allowed workspaces; opening a terminal does not add one automatically.
The browser chooser offers repositories the host allows. This illustration uses a fixed fictional choice.
A browser can view an existing mirrored terminal or start its own session in an allowed repository.
A mirrored session belongs to its terminal. A browser-started session can outlive its tab. Neither session is running in this demonstration.
Founder / company
Conor "Lone Pine" Sullivan is a game developer working under the Lone Pine Games brand.
SLAM grew out of using AI in game development, where the human still needs to direct what to build. Hybrid development remains the goal; it is still a work in progress.
Back to home