3 agents running · run r-9d41 passingworkspace brain · 1.2M chunks indexedsession #a4f2 · synced across 3 devices
Agent workspace · v2.4

Your terminal,
with a fleet behind it.

Jarvis orchestrates workers, reviews their diffs and merges the result — while you watch every command travel the graph in real time.

Download for macOS Read the docs
Apple Silicon · macOS 13+ · free while in beta
on the name

A name you have to live up to.

J.A.R.V.I.S.onlineAB
dex-pair-intelligencecandle-retentionrun/drift-fix/wsnAnthropic · opus 4.84
1drift-fixdex-pair2login-bugadmin3vault-rotatesecrets+
@code-review (go-code-reviewer) — the diff touches a hot path
$background
✓ 48 passed · cov 86%
Read(head-of-architecture.prompt.md)
1---
2name: head-of-architecture
3description: "Deep architectural thinking, ArchiMate / C4 modeling, microservices and async patterns."
4spec-version: 1.6.0
5model: opus
6tools: [read, grep, tree, adr]
7escalation: design-council
8memory: shared/architecture.md
… +175 lines (click to expand)
Bash(ls ~/.jarvis-runs/brains/ws-5dfa/agents/head-of-architecture/)
IDENTITY.md
head-of-architecture.prompt.md
WORKING
MR !138merge request SYNCEDgitlab · main Spec readload canonical spec Extract IDsAST · canonical IDs Detect driftstructural set-diff Residual proseLLM prose drift Reviewconformance §6 Syncrender + publish SPEC-INDEXcanonical spec · $0.002 ASTextractor · $0.001 DRIFT-ENGINEdeterministic · free OPUSprose drift · LLM · $0.090 GITLABprovider · sync · $0.005
EST$0.000/ run
OPUS drives 96% · 6 steps · 5 models
// models
Your accounts.
Jarvis picks the model.
calls by hour · last 7 days
quietbusy
cost · august
$384total this month · 41 200 calls
plans · fixed
$200/ mo
3 seats · Max, Team, Pro
paid whether you call or not
api keys · metered
$184of $400 cap
billed per token · overflow only
calls served
subscription seats 71% · $0 marginalapi keys 29% · $184
effective cost $9.32 per 1 000 calls · seats absorb 71% at no extra charge
agents · rosterevery run writes back
HOHead of Architecturev3.2.0
drift engine · root-cause · bound to Claude Opus 4
Knowledge map — track record per domain
Drift detection0.94
50 runs · 94% accepted · weights this agent in conflicts
Go architecture0.91
49 runs · 91% accepted
Conformance §60.83
46 runs · 83% accepted
Prompt history — GitLab · 4 tags
v3.2.0headauto-train2d ago
Learned the §13.U tie-break resolves to lower tier — encoded as a hard rule after the prose-drift conflict.
v3.1.0auto-train9d ago
Added multi-tenant resolver knowledge from the SD-6 epic.
v3.0.0manual3w ago
Re-trained on the v0.5.1 spec refactor; conformance reasoning rebuilt.
sync · session #a4f2state lives in the cloud — not on this machine
jarvis> run checkout tests
├ payments-api … pass · 1.2s
├ cart-sync … pass · 0.8s
└ e2e-checkout … pass · 3m 12s
run complete · artifacts saved · nothing left on disk
Session state moved to the cloudnothing left on this machine — pick it up anywhere● SYNCED
#a4f2cloud session● synced
14:31
J.A.R.V.I.S.LIVE
session #a4f2 · checkout-flow
continued from MacBook · just now
jarvis> status
├ payments-api · pass
├ cart-sync · pass
└ e2e-checkout · pass · 3m 12s
ready to deploy → staging
✓ deployed · staging · 14:32
state synced back to all devices
Workspace memory

Every human and agent works through the same memory.

The workspace is the source of truth — Jarvis understands it, and any agent works through it.
Coding agentClaude Code · Codex · Cursor
Human terminalcommands · sessions · control
JARVISworkspace brainmemory · context
orchestration
CI / Reviewchecks · tests · evidence
Bring any agentIDE · CLI · bots · custom
Retrieval

Recall in 38 ms, not minutes.

The index is built at write time — recall is a lookup, not a search.
4.2 s
grep repo + logs
2.6 s
IDE full-text
900 ms
generic RAG
38 ms
jarvis recall
time to answer · lower is better · 110× faster
Indexed at write timeEvery trace, diff and decision is chunked and embedded the moment it happens — nothing to crawl later.
Workspace-scoped graphMemory is linked: a run knows its diff, the diff knows its review, the review knows who decided.
Answers with provenanceEvery recall cites its sources — run id, MR, date — so trust is one click away.
jarvis> recall why checkout e2e flaked
→ race in cart-sync · run r-9d41 · fixed by MR !138 · decided Aug 12
Execution

Two places to run. One interface.

Every task lands in a real session — isolated in the cloud or right on your Mac. Same commands, same graph, same memory.
cloud sessions
Isolated, restart-proof, disposable.

Each session gets its own workspace, git worktree and lifecycle on the cluster — spun up per task, surviving restarts, thrown away clean.

  • uid-isolated per session
  • transcripts and artifacts in object storage
  • close the lid — the run continues
run on this mac
Your disk, your toolchain, no tunnel inward.

Pair a machine and sessions run locally on its own disk. Nothing is copied to the cloud, no VPN, no inbound port — the machine dials out.

  • paths the daemon may touch are yours to pick
  • local engines and shells as they are
  • tunnel opens outward, revocable in one click
sessions close honestly
A session ends as merged work, not a stale tab.

On close Jarvis commits, opens a merge request per touched repository, waits for green, merges into dev and publishes what it learned into workspace memory.

  • one MR per repository, always
  • green pipeline merges itself
  • decisions and traces land in memory
privacy gate
Secrets never reach the store.

Everything published to memory passes a gate: credentials, tokens and high-entropy strings are dropped before they are written, not redacted after.

  • filtered at write time, physically absent
  • hybrid recall stays keyword + vector
  • every answer cites its source
KubernetesGitLab CI/CDArgoCD GitOpsS3 artifactsBM25 + vectorsCodices
One command

The fleet is one line away.

Signed, notarized, free while in beta.
$brew install --cask jarvis
macOS · Apple Siliconv0.2.21 · dev build · 38 MBsha256 9d41…c77a
Version history0.2.21 is a dev build — expect rough edges.
0.2.21currentdev buildResponsive input across 20–30 sessions; only the active raw terminal is mounted, recent tabs stay warm, and large session gaps restore from a bounded tail without losing background jobs.Aug 14, 2026.dmg ↓ 0.2.20Session tracker truth, brain transfer status, and durable replay without partial transcript flashes.Aug 13, 2026.dmg ↓ 0.2.18Session picker restored beside its plus button, with outside clicks passed through in one action.Aug 13, 2026.dmg ↓ 0.2.17Workspace brain recall with provenance chips; faster cold start on large workspaces.Aug 12, 2026.dmg ↓ 0.2.16Flow cost breakdown per step; endpoint cards for MR and sync targets.Aug 11, 2026.dmg ↓