The product
Stop building features.
Start building teams.
Build autonomous spec-writing, code-churning, self-auditing teams, with full decision provenance, that learn and act on your canonical values. Manage the fleet and build overnight.
$39 a month, 14-day trial. Your own Claude login, no API key.
Overnight23:04 → 07:12
Spec3specs written
Build6pull requests opened
Audit5findings judged, 4 refuted
One mark is thirty minutes.
Waiting for you1 contested rule
One agent is a tool. Five is a team.
Scale the fleet without scaling what you have to watch.
9agents at peak2asked for you
22:0000:0002:0004:0006:00
It learns while you work.
Sessions distill into governed rules through six passes, autonomously, in the background. Every drawing below is a stop on this line.
pipelinesix passesone line
ingest
session 118a session lands, raw
extract
14 claimsdiscrete, each with its source
reconcile
1 conflictrefereed by rank, kept on record
curate
2 proposalscanon candidates, decaying unless said again
gate
1 ratifiedby you, or autonomously under your rules
render
canon, byte-stablesame input, same bytes, full provenance
ingestsession 118 a session lands, raw
extract14 claims discrete, each with its source
reconcile1 conflict refereed by rank, kept on record
curate2 proposals canon candidates, decaying unless said again
gate1 ratified by you, or autonomously under your rules
rendercanon, byte-stable same input, same bytes, full provenance
Claim a warm tree in two seconds.
Every agent gets its own branch and its own checkout. Nothing collides.
- Two agents cannot take the same tree.
- Nothing runs in the checkout you have open.
main8f3c2a16 worktrees
- pool-1feat/rate-limitrunning4m 12s
- pool-2feat/upload-resumeci green6m 30s
- pool-3feat/retry-backoffasking you11m 05s
- pool-4feat/session-indexrunning1m 48s
- pool-5chore/pool-warmuprunning2m 04s
- pool-6freefix/order-id-searchwarmclaimed2s to claimin 2s
Fleet tokens today664k
feat/rate-limit160k
fix/session-flake443k
feat/session-index61k
chore/autonomy0
Every session, live.
Read what each one is doing, and type into any of them mid-run.
- Closing the window does not stop the run.
- Park an idle session and resume it later.
sessionfix/session-flakelive, 11m
agent pool-2
turn 6
- Ranvitest session-surface11:42
- Readsrc/hooks/use-turn-paint.ts11:43
- Editeddebounce paint until the turn settles11:45
- Ranvitest session-surface: 42 passed11:47
>hold the flake fix, ship the paint fix first
The work writes the board.
Agents open, advance, and close their own tickets as they work.
- A commit cannot land untracked.
- Stages only ever move forward.
boardsession-flakeforward only
identified
wi/session-flakeopened from the review finding; no branch exists yet
in progress
fix/session-flakethe branch is born tracked; a commit cannot land without the item
pr open
pr 118the spine stamps the PR number onto the item, not a human
ci
checks 6 of 6the item holds while the checks run; green is not yet merged
merged
b4c7e21 · branch gonethe squash sha lands on the record; the lanes show this same card
identifiedwi/session-flake opened from the review finding; no branch exists yet
in progressfix/session-flake the branch is born tracked; a commit cannot land without the item
pr openpr 118 the spine stamps the PR number onto the item, not a human
cichecks 6 of 6 the item holds while the checks run; green is not yet merged
mergedb4c7e21 · branch gone the squash sha lands on the record; the lanes show this same card
backward advance · refusedthe engine refuses it; a stage is a fact, not a label someone can drag back
Teach the fleet once.
Finished sessions distil into rules. Every later run is judged against them.
- Two rules that contradict cannot both survive.
- Only disputed rules reach you. The rest just hold.
canonretry-capdecision · half-life 120d
d0
Distilled from a finished session at confidence 0.92.
d2
The gate ratifies it, uncontested. You were never asked.
d8+
Every later run is judged against it while it holds.
d40
You say it again in a review. The clock resets to 0.92.
d47
Its unrepeated twin crosses the 0.7 floor and quietly stops being promoted.
Know why an agent did that.
Every run links back to the rule it followed, and whoever set it.
- The edges are written, not guessed.
- Retract a source and everything it taught goes with it.
askWhy did run 121 cap its retries at three?
- run121
- judged againstruleretry-cap
- 5 later runs
- 3 merged claims
- 1 rule it replaced
- ratified fromclaim3 retries, not 5
- extracted fromsessionfix/flaky-auth
- written byauthoryou, 03-11
Your word outranks theirs.
Trust rules referee the collision, and the loser stays on the record.
- A model may flag the clash; it never picks the winner.
reconcileone subjectrank decides
operatorretry cap: 3you, in review
layertie
sourcedecides
recencynot read
confidencenot read
extractedretry cap: 5raw session 118
retry cap: 5conflicts_with →retry cap: 3kept, not deleted
Undo what a bad session taught.
The graph knows every lesson's source, so one retraction empties exactly that lineage from canon.
- The cascade is atomic, and it is audited.
taintsession 47one pass
e2e is flakyderived_fromsession 47
rule · skip e2ederived_frome2e is flaky
cap retries at 3derived_fromsession 12
retract session 47its claim and the rule it taught leave canon together; session 12's claim does not move
Max the autonomy. The floor holds.
Two settings, gated and full. The difference is who dispatches an agent and who merges a green pull request.
- Tighten one run without a commit.
gateautonomy3 of 10 always held
an unrecognised value resolves to gated
| what the fleet may do | gated | full | |
|---|---|---|---|
| open a work item and a branch | automatic | automatic | |
| dispatch an agent into a worktree | recorded, not run | automatic | |
| merge a pull request that is green | waits for you | automatic | |
| run the loop overnight, unattended | waits for you | automatic | |
| no setting reaches below | gated | full | |
| widen its own autonomy settingthe setting is a committed file | refused | refused | |
| resolve a review thread it wrotea second agent has to close it | refused | refused | |
| keep merging without a reviewthe loop stops itself and says so | halts the loop | halts the loop |
No finding is real until it reproduces.
Agent reviewers cry wolf; this one has to show you the trigger firing before a word of it is believed.
- Nobody closes their own fix.
- A refuted finding keeps the quote that killed it.
reviewone findingrefute-by-default
filed · refuted
a trigger runs
reproduces · 9 of 200it is real
promoteda fix lands
fix on pr 118not the author
cleared by pool-4filed · refuted
a trigger runs
reproduces · 9 of 200
it is real
promoted
a fix lands
fix on pr 118
not the author
cleared by pool-4
never reproducesdies as refutedand keeps the quote that killed it, on the record
defaultevery finding is born refuted; belief is the state that has to be earned
promotiononly a reproducible trigger promotes; an argument, however good, does not
clearancean agent that did not write the fix closes the thread; nobody grades their own work
Corrections append. Nothing overwrites.
Every action carries its actor, and names whatever it replaced.
- Your own wrong calls stay in the record too.
trailone recordseq-ordered · append-only
seq14372assigned by the store, never by the writer; the log’s total order
kindatom.supersededwhat happened, from a closed vocabulary
actoroperatorwho did it, on every record, yours included
at2026-08-14T09:41:55Zstamped once at append, never revised
atom_ids[retry-cap.e41, retry-cap.9b2]the elder it retired and the twin that replaced it
retry-cap.9b2 · activesupersedesretry-cap.e41 · supersededthe correction points backward; the elder is never edited and never deleted
appendthe only verb the log has; every action lands as a new record
no overwriteno write edits a recorded event; a fix is a new record naming the old one
no deletesuperseded means dimmed and kept, so the wrong call stays readable beside its correction
Your token usage, line by line.
Attribute every token to the run, branch, and model that spent it.
- A held branch spends nothing.
- A second opinion costs a run.
usage560 runs · 23.1M tokens this month
session-flake merged 10:26443k
- sonnetReproduced the flake, 9 of 200 runs64%284k
- opusWrote the fix, 3 files22%96k
- opusVerified it independently9%41k
- haikuBoard, trail and canon writes5%22k
today, all projects4 projects · 1 held
- api9 runs1.9M
- chore/autonomy0 runs0held by the gate
- web6 runs1.1M
- infra3 runs480k
- docs1 run20k
- today19 runs3.5M
by model, this month
runs
tokens
- opus
- sonnet
- haiku
Your code never reaches us.
The whole graph is one file on your disk, in an open format.
- Point a packet monitor at the machine and read the zero.
localone filetyped edges
findingderived_fromsession
rule · retry-capderived_fromfinding
rule · retry-capsupersedesretry-cap v1
sessiontrackswi/session-flake
atomsfive drawn: a session, a finding, a rule and its predecessor, a work item, each a row carrying its author and its source
edgestyped at write time: a rule knows which finding taught it and which older rule it replaced
yoursdelete the file and the record is gone everywhere, because there is no second copy on someone else's machine
Seventeen views, one window.
Sessions, worktrees, runs, spend and canon, with nothing else to open.
runs
orchestrate.so
OrchestrateHomeSessionsLiveActivityWorktreesREPLRunsDaemonSpend
GovernBoardProjectsInboxGateDeliberationTimeline
KnowledgeMemoriesCanon
17 viewsHomeSessionsLiveActivityWorktreesREPLRunsDaemonSpendBoardProjectsInboxGateDeliberationTimelineMemoriesCanon
Runsshowing 50 of 4,118
workingreconcile2m ago118 claims
workinggate9m ago3 ratified, 1 held
endeddispatch14m agofeat/retry-backoff
endedextract22m ago41 claims from 6 sessions
heldrender31m agoheld on conflict
gatedliveup 2d 4h · pid 26928 · 3/4 agents
Build. Autonomous. Teams.
With everything Claude Code should have had.
$39 a month, 14-day trial. Cancel from the app.