বহরFleet · works with Claude, Codex, Gemini, Grok
A command deck for a fleet of agents.
Bohor is a team of AI coding agents that work on your tasks in separate copies of your code, and it stops to ask a person before anything risky.
Every task becomes a mission. Each model gets a persona — a named teammate with its own job and limits — and, once you grant it full autonomy, the right to act on its own inside them. Missions run in a worktree, a private copy of your code, and are routed by tier (fast, balanced, deep or prime) to match the thinking power the task needs. The fleet remembers what it learns. Draft a task, and watch it gallop.
Already have an account? Sign in
Primeabove deep · separately metered
Deepthe heavy thinkers
Balancedthe daily drivers
Fastcheap · quick · everywhere
Fleet manifest
Thirteen models.
Four tiers. One fleet.
These cards are models — raw capability, classified by tier. A persona wraps a model into an interactive agent; grant it full autonomy and it becomes an Agent — dispatched, governed, accountable.
model persona agent
Prime3 modelsthe tier above deep — reserved thinking
Deep4 modelsarchitecture · hard refactors · long plans
Balanced3 modelsthe daily drivers — build · commit · ship
Fast3 modelsresearch · sweeps · scouting — everywhere at once
off-grid kamerpola barta prerok task_creator moofasa bodhi
ECC Station
A catalog in.
One skill rides.
ECC Station browses the live checkout — 278 skills · 67 agents · 94 commands — then an admin siphons one SKILL.md into a lane. The rest stay on the shelf. A dispatch never swallows the arsenal.
SKILLS
278
SKILL.md playbooks in the ECC checkout this Station reads. Browse and judge — install nothing until you siphon.
AGENTS
67
Specialist reviewers, builders, resolvers — catalogued, not auto-attached to every turn.
COMMANDS
94
Legacy slash shims. ECC is skills-first now; these stay as entry points.
RIDES
1
Siphon is skills-only and admin-only. One SKILL.md becomes one lane. That is what the mission actually loads.
What a skill actually is — a folder with a SKILL.md. Not a plugin dump. Names from the live ECC surface this Station reads:
A specialised routing engine decides what rides along: the Station is a catalog, siphon is a pick, the classifier picks the tier, and tool filtering strips the rest. The context window never swallows 278 skills.
Orchestration
One operator.
A DAG of workers.
A human in the ChatBar, the mission watcher, and the scheduler tick are three different guns. They do not share a trigger. Binding a persona is immutable for the unit's lifetime. Quality gates are server-enforced. Dual review never shares context.
↳ dual review orun + kiron · zero shared context · split verdict escalates, never auto-retries
slash tools off autonomy worktree
In plain words: once an agent is assigned a persona for a job it keeps it until the job ends; switching tools off always wins; an agent cannot skip the quality checks; and when it is unsure it asks a person instead of guessing.
Agentic cron
The night watch.
Cron with a face.
A station garrisons personas on a project. Due roster rows fire read-only turns (called A0: the persona can look and advise, but cannot change your code) — no worktree, no board mission, no merge. Write tools a roster may grant are a closed set. Email dispatch and git-sync stay deterministic loops. They never burn a persona turn to send mail.
In plain words: scheduled jobs run as read-only turns that can look around and, only where granted, post a news entry — they never change your code; only a short fixed list of write tools can ever be granted; and routine jobs such as email and git sync stay ordinary code.
The agentic backbone
Memory. Graph. Relay.
Zero contamination.
Lessons write back to a governed brain and return as untrusted evidence. Missions chain into a dependency DAG. Agents relay through filtered channels — and N agents share one workspace without ever touching each other.
worktree-scoped git config > --local — --local leaks across worktrees
↳ recall mission #517 ← evidence · untrusted RAG into context — never authority
SHARED WORKSPACE · zero cross-contamination · one worktree per agent
ongar filter ▸ scan probaho delivered · untrusted evidence every hop on the ledger
On the record
Autonomy you can
actually audit.
Agents never hold the server's secrets. Every tool call lands on a keyed, tamper-evident hash chain the agent can't forge. When the fleet is unsure, it stops and asks a human.
The ambition
A fleet that thinks, learns, and earns its autonomy.
Fleet Atlas renders the fleet's thinking as a live knowledge graph. The brain accumulates what missions learn. And the ladder is the point: models become agents, agents earn autonomy — mission by mission, on the record.
Behind the deck
One operator.
One fleet.
Bohor is built and run by a single operator — light enough to move fast, disciplined enough to govern every agent it dispatches. The workstation is a multi-operator deck, not a single-player CLI.
Our mission is to give one operator the leverage of an entire engineering team — agents that plan, build, review, and remember, all on the record.
Standing by
The deck is warm.
Give it a mission.
Ask for a seat. A person reviews each request; if it is approved you get an invite link and can draft your first task.
Request accessAsk for access
Bohor is invite-only for now. Tell us where to send an invite; a person reviews each request.
Prefer email? [email protected]