The real work lives between your agents — remembering what changed, handing it off, knowing when to stop. That wiring has always been on you. BitCadence is the loop manager that carries it: one governed job board with human approval gates, an immutable audit trail, and Drumline — a shared memory every agent reads and writes.
Most agent fleets have amnesia: every task starts from zero. BitCadence gives the whole mesh one collective memory.
Completed jobs are auto-distilled into recallable handoffs. Agents record
facts, decisions, and lessons deliberately with mco_remember.
Before any worker executes a lease, the most relevant entries are injected
straight into its prompt.
Drumline isn't an embeddings bill or a vector-DB subscription. It's a deterministic, auditable scoring substrate that ships inside the product — and persists to an embedded SQLite store on your machine. The free edition has the full memory. That's the point.
Built from validated enterprise demand: the controls IT leaders asked for before they'd let agents touch production.
A Score task can commit real code to an isolated, disposable branch —
never main, never pushed by the automation itself,
blocked by a hardcoded denylist and path allowlist, not just operator
discipline. Every commit needs a signed, human-issued grant scoped to that
exact run; an agent token can never satisfy the human-auth check, no matter
its permissions. On 2026-09-18 this shipped its first real commit: one agent
wrote it, a second independently reviewed it, a human authorized it.
Flag any job — or force-gate entire roles like servicenow — and it pauses
at needs_approval until a human decides. From the console, the CLI, or MCP.
Every mutation appends to agent_job_events; UPDATE and DELETE are rejected
at the storage layer. Inspect any job's full history with mco audit.
Mission-control overview, job board with audit drawer, approvals inbox,
visual workflow builder, agent fleet presence. One HTML file at /console.
Failed jobs retry on a budget, then auto-escalate to a human role — or open a ServiceNow incident through the escalation bridge.
One env var pauses all intake and leasing mesh-wide. In-flight work reports in; humans can still approve and audit.
Multi-agent pipelines in YAML — per-step approval gates, retry budgets, and escalation roles. Or draw them in the visual builder.
The board as a diagram at /flow, where every edge is a real
depends_on gate the runtime enforces — not decoration. Click a node for its
audit trail, approve or cancel inline, or drag to author a workflow and export the YAML.
Cron and interval schedules that run through the governed board. Loops must declare how they end — a count, a deadline, or "until the queue is clear" — or they're refused at parse time. Every scheduled job is stamped with what created it, so the audit trail answers a question plain cron can't.
API keys, tokens, and passwords go to an AES-256-GCM store, never silently to plaintext
.env. Auto-provisioned on Windows; a deliberate opt-out anywhere else.
Ingest ServiceNow incidents and Dynatrace problems as jobs; act back with auditable, approver-gated platform actions.
No Supabase creds? An embedded SQLite store takes over automatically —
jobs, audit, registry, and Drumline persist in ~/.mco/local.db.
The expensive agent should not have to think deeply about every small routing decision. And a fast guess should never become permission to act.
Why it exists. Agent fleets waste time and context deciding who should do a task, which evidence is relevant, or whether an alert is noise. Jev, built on TypeSafe System One, returns typed judgments and confidence for bounded questions. BitCadence can use that advice to prepare a better task packet and choose an appropriate model or worker.
Two speeds, one chain of responsibility. Jev handles a small semantic judgment; Claude, Codex, Gemini, and other workers do deeper work. Deterministic code still owns budgets, eligibility, leases, approvals, and effects. A Jev receipt records what was asked and returned; it cannot grant itself authority.
In this release, Codex and Claude routing are advisory, while Drumline, watchdog, and notification checks add annotations only. Jev is optional and off by default. BitCadence still runs locally without a TypeSafe account, key, or network call.
A separate, owner-approved VIA path may let Jev choose among schedule candidates extracted from retained parish evidence. Code rechecks the chosen candidate against that evidence and applies deterministic fallbacks; this bounded selection does not give Jev authority to create evidence, bypass review or budget rules, or make other changes. That VIA implementation is separate from this release.
/api/jev with mode: shadow, model: jev-latest, and the key. The key goes into BitCadence's encrypted vault; never place it in this website or a browser script.POST /api/jev/test to check model discovery. Then inspect Jev decision receipts and fallback behavior before considering a pinned model or another mode.Each agent gets a hashed bearer token and a role. Any agent may send work to any role; only the addressee can lease it. Like dropping mail in a box.
From the console, the CLI, MCP tools inside Claude Desktop, a YAML workflow, or a ServiceNow webhook. Dependencies chain; gated jobs wait for a human.
Workers lease atomically, run through the right vendor CLI, report back. Every outcome is audited, and distilled into Drumline for the next job.
Same engine everywhere. The free edition is not a demo — it's the full product on your own machine.
No cloud account. No API keys. No database to provision.
curl -sSf https://bitcadence.ai/install.sh | bash
macOS & Linux · Windows instructions below
One command installs everything and walks you through setup:
# Works on macOS and any Linux distro curl -sSf https://bitcadence.ai/install.sh | bash # Clones the repo, finds/installs Python, creates the venv, # generates your access token, adds 'mco' to your PATH, # then asks: demo mode or connect now. # Or if you already cloned: bash scripts/install.sh
After install: mco serve starts the gateway · mco setup changes any setting
Double-click installer — no terminal needed:
# 1. Download the ZIP from GitHub and extract it anywhere # (e.g. C:\BitCadence) # 2. Double-click: install.bat # Finds (or installs) Python, creates .venv, generates your # access token, copies a BitCadence shortcut to the Desktop, # then asks: demo or connect now. # Your browser opens the console automatically.
Or run headless: powershell -ExecutionPolicy Bypass -File scripts\install.ps1 -NoPrompt
For developers who prefer explicit control:
# Clone + pip editable install git clone https://github.com/mastalink/BitCadence pip install -e BitCadence mco setup --guided # configure in 60 seconds mco serve # console at 127.0.0.1:18789/console # Docker (team / cloud deploy): docker compose up # see docs/DEPLOYMENT.md
Workers join from any machine with mco listen --role codex
Those are frameworks for building agents in one codebase. BitCadence coordinates agents you already have — Claude Desktop, Codex CLI, Gemini, custom workers — across machines and vendors, with the governance layer (approvals, audit, escalation) those frameworks leave as an exercise for the reader.
Yes — that's a product decision, not an accident. Shared memory is the heart of the system, so it works fully offline against the embedded SQLite store. Cloud databases add multi-machine and multi-tenant reach, not features.
Every job mutation appends a row to agent_job_events. UPDATE and DELETE
are rejected at the storage layer — by a database trigger in cloud deployments, and
by the storage engine in local mode. History can be read, never rewritten.
Yes. Workers make outbound-only connections to the gateway, so they run happily inside customer networks or behind NAT — the control-plane/data-plane split enterprises expect from credential-holding tools.
Most shipped BitCadence integrations provide advisory routing or annotations. A separately approved VIA Lorain path may use Jev to select among evidence-backed schedule candidates, with code revalidating the selection and enforcing fallback rules. That path is separate from this release. Jev cannot create evidence, grant permissions, approve checkpoints, or control unrelated effects; unavailable or invalid responses follow deterministic fallback.
MIT. Self-host it, fork it, ship it inside your company. The enterprise pilot program is for teams who want connectors, multi-tenancy, and a direct line to us.