EXECUTION • SCHEDULING • RECOVERY
Durable background execution for AI agents.
Run long jobs, workflows and schedules without losing work when agents, workers or machines restart — entirely on your machine. Cronus runs.
One line. No Git, no Rust, no admin rights. The installer pins the release public key and the SHA-256 of the bootstrap; the bootstrap verifies the signed catalog and every archive (hash + Ed25519) before anything runs.
Windows x86_64 — PowerShell:
irm https://github.com/titans-tools/titans-platform/releases/latest/download/Install-Cronus.ps1 | iexLinux x86_64:
curl -fsSL https://github.com/titans-tools/titans-platform/releases/latest/download/install-cronus.sh | shAlready have any Titan? Then it is just:
titans install cronusPrefer a manual download? Signed archives per platform are on the
Cronus releases page
(bin/cronus inside; SHA-256 and signature in signed-distribution-catalog.json
on the platform distribution release).
Shared platform tools and the signed catalog live in the titans-platform repository. macOS and ARM64 are
not part of this release.
Kill a worker mid-job and watch the work survive:
submit job
↓
RUNNING ━━━━━━━━━░░░░░░░ 46%
↓ (worker killed)
RECOVERING — lease expired, claim fenced
↓
RUNNING ━━━━━━━━━━━━━░░░ 72% (from the last checkpoint)
↓
COMPLETED ✓ result archived to Atlas
| The problem | How Cronus answers |
|---|---|
| The agent must not block on long work | Submit + poll |
| A worker died mid-job | Lease recovery — never double-claimed |
| Work failed | Retries with backoff, then a dead-letter queue |
| The flow has many steps | Durable DAG workflows |
| It must run later, or every day | Scheduling (@every / cron-like) |
| A step needs a human decision | Approval gates |
Cronus is the execution Titan. Long or heavy work — extraction, validation, model downloads, tool installs, renders — is submitted to Cronus instead of blocking an agent inline. Jobs run in lanes with durable at-least-once execution, idempotent claim and completion, checkpoints, lease recovery, dead-letter handling and schedules, and leave a durable trail. Results are archived to Atlas through a durable outbox.
Cronus makes asynchronous and autonomous workloads durable: work survives failures, processes can restart, and agents stay continuously operational — without turning the execution layer into a knowledge store (that is Atlas). It runs as a real background service (Windows Service / user daemon) and is operated from a built-in Workbench in your browser: what is queued, running, stuck, dead-lettered, scheduled, and who is doing the work.
Core capabilities: job queues · DAG workflows · workers · scheduling · watchers · supervised daemons · retries · leases · backpressure · checkpoints · failure recovery
| Capability | Tool | Result |
|---|---|---|
| Submit durable work | cronus_job_submit {definition:{kind, owner, payload_ref, tags}} |
kinds: knowledge_extraction, maintain, quality_evaluation, tool_install, model_download, document_render, video/audio/image, … |
| Track it | cronus_job_get · cronus_job_list · cronus_job_progress · cronus_job_artifacts |
status, checkpoints, progress %, output refs |
| Control it | cronus_job_cancel, cronus_job_require_approval · cronus_job_resolve_approval |
governed cancel + human/agent approval gates |
| Schedule recurring work | cronus_schedule_create · cronus_schedule_list · cronus_schedule_pause |
@every / cron-like presets |
| Workflows | cronus_workflow_* |
multi-step runs with durable state |
| Recover | cronus_recover, cronus_dead_letters · cronus_dead_letter_replay |
a killed worker's job is recovered, never double-claimed |
| Lanes & fairness | queue tools | dedicated heavy/simulation lane, backpressure, priority, DLQ |
Reliability guarantees: durable at-least-once execution, atomic/idempotent claim and completion, stale-claim fencing, per-task panic isolation (one job crashing never takes down the worker), lease recovery, durable archival delivery through the outbox. Exactly-once external effects require a connector that persists the Cronus idempotency key.
cronus statuscronus call cronus cronus_job_submit '{"definition":{"kind":"maintain","owner":"me","payload_ref":"atlas://acme/alpha","tags":["demo"]}}'cronus call cronus cronus_job_list '{}'Open the Workbench: http://127.0.0.1:7743/workbench.
- Cronus speaks MCP over stdio; the installer wires it into the shared
titansMCP server of the AI clients present on the machine. - Payloads travel as canonical references (
atlas://), never inline bytes. - Submit-and-poll: submit a job, then poll
cronus_job_get; don't block. - Skill:
skills/cronus-durable-jobs.mdteaches an agent the whole product — job lifecycle, lanes, the fenced worker protocol, schedules, watchers, triggers, workflows, daemons and the full tool index. Alsotitans skills show cronus-durable-jobs. - REST (
127.0.0.1:7743) and gRPC are loopback-only by design.
Cronus publishes an operational floor, not an unbounded throughput claim: the release gate requires ≥ 25 submits/s (p95 ≤ 100 ms) and ≥ 10 terminal transitions/s (p95 ≤ 200 ms) over 1,000 durable jobs. Lane and worker concurrency are explicit, configurable backpressure controls; exceeding a configured capacity queues work instead of weakening claim fencing.
Two independently signed channels — a public discovery catalog and a distribution catalog — drive updates. Nothing changes without explicit confirmation, and a version can be pinned.
titans versions checkThis release is free. Capabilities and versions that become paid are published in the signed public catalog; rights you acquire for a version stay with that version (see the EULA, §1.3).
Quickstart · Skill for AI agents · Security model · Using from AI agents (MCP) · Changelog · Support
Report vulnerabilities privately — see SECURITY.md.
Proprietary. Use is governed by the Titans End User License Agreement. Third-party components are listed in the SBOM inside every release archive. This repository hosts the official product page; the source code is not published.