AI job orchestrator
Posse queues your work, plans the jobs, and runs a crew of AI agents across all your repos at once — work branches, merge contention, secrets, and token spend handled automatically.
Direct APK — allow "install from unknown sources" on your device.
What's in the outfit
Everything an unattended crew of agents needs to do real work on real codebases, and nothing that lets them hurt you.
A lease-based scheduler keeps the herd moving: work items are decomposed into jobs, ordered by their dependencies, and routed to the right rider — with expired leases recovered and deadlocks cancelled automatically. Scale the crew up without babysitting the queue.
Scoped tools won't touch credentials, so secrets never enter a prompt at all — with an independent secret scan as a second fence. Models can't leak what they never saw.
Every symbol indexed, every relationship charted, a semantic layer over it all. Agents traverse the graph instead of reading file after file — the whole codebase understood through a fraction of the context.
Because Atlas scopes the context, agents reason over what's relevant instead of wading through the whole repo — better answers and a smaller bill from the same model.
Teach the posse your conventions: package custom skills and tools once and every agent on every repo rides with them.
Point the whole crew at one codebase — even the same files — without anyone stepping on anyone's toes. Every job gets its own git work branch, and file locks with job ordering decide who rides first when paths overlap. Git keeps the work isolated, locks keep it ordered — merges land with zero contention.
Set your own spend budgets and context caps — enforced like every other rule, in the tools. And every prompt, output, and cost is recorded per job: know what each agent spent, on which model, doing what, down to the token.
Handoffs never dump context into an agent's lap. Prior research, plans, and work products live on the backend, filed by content hash; each agent is handed a curated map of only what it needs — cutting through the noise and keeping it focused on its task. Anything deeper sits a semantic hop away, expanded on demand.
What an agent learns is pinned to the exact functions and files it describes, and resurfaces precisely where it applies. When the code underneath changes, the memory's confidence fades — it stops being offered instead of lying to the next agent.
Route jobs across Claude, OpenAI, Codex, Grok, and Copilot — or keep work on models running on your own hardware. Pick the right horse for each job; swap providers without touching your workflow.
Atlas
Atlas tokenizes and indexes every symbol in your codebase, then charts the relational map between them all — who calls what, what depends on what — with a semantic layer over the top, so the whole graph is searchable by meaning.
An agent riding Atlas doesn't hunt through files. It traverses the map — symbol, to caller, to dependency — and understands the codebase through a fraction of the context. That's not just cheaper. The grep-and-read pattern floods a context window with code that doesn't matter, and that rot is exactly what drags agents off task. Atlas never lets it in the door.
Context loaded for the same job.
Deterministic guardrails
Posse doesn't secure agents by writing sterner instructions. The rules are compiled into the tools themselves — so safety never depends on a model deciding to behave.
Agents never touch their stock tools. Posse swaps in its own tool set, where every call is permission-checked deterministically — and every job carries a hard scope, the files it's allowed to touch, enforced at the tool boundary. An agent that ignores its instructions still can't act outside its grant. Work is also carried out on independent work branches that refuse to merge out-of-scope content, providing a second layer of protection.
The scoped tools flatly refuse to touch secrets — API keys, tokens, and connection strings never enter a prompt in the first place. Behind that boundary stands an independent secret scan, a second fence that catches anything secret-shaped before it can slip out. And every assembled prompt is logged, so you can verify exactly what the models were shown.
Design your own approval pipeline — wave the routine work through, gate the risky paths, pause for a human wherever you choose. And no decision is final: send a change back with notes and the agent fixes its own work, or roll it back clean and ride on.
Bossy
Bossy rides above every posse on your machine — each repo's agents, queues, and live jobs at a glance, with the tab rolled up in real time: tokens, hours, dollars.
Fleet overview
screenshot coming
Jobs on the trail
screenshot coming
Tokens & spend
screenshot coming
Frontier when you want the best, local when it shouldn't leave the building.
Saddle up
Queue work, watch the trail, and approve review gates from the iPhone or Android app — your agents keep riding while you're away from the desk.
Posse
