Frequently asked questions

Ogun is a local-first developer platform for data acquisition, data engineering and AI-agent benchmarks, made by Verne. It turns authenticated, messy web and filesystem sources into verified, reproducible datasets on a schedule, and it can run benchmark tasks for data and terminal agents in isolation. It is built on the Zeus download engine.

AI, LLM and data engineers who build datasets and pipelines, ML researchers who need provenance, and people who evaluate coding and data agents. If you want everyday downloads for your AI assistant instead, use Zeus.

As a pre-alpha. Release 0.0.1a has Ogun Studio for Windows and Linux and the ogun and ogund binaries, and a hosted hub runs Studio on the real daemon with sample data (see Get started). The runtime passed its phase O1 exit test on Linux in October 2026: a partitioned backfill and a stream window survive kill -9 of the daemon and resume only the unfinished work. Code operators, remote approvals and most connectors are not ready; expect breaking changes. See the roadmap.

Both are Verne products built on the same engine library, zeus-core. Zeus is the session-aware download and file layer for AI agents; Ogun is the developer platform. Ogun embeds the engine directly and reaches browser-session sources only through the Zeus daemon, as a service client. That bridge is built and tested against a fake daemon.

Ogun is free and open source under Apache-2.0. A hub for teams (ogun-server: OIDC sign-in, workspaces, roles, edge sync) is built and runs the hosted pre-alpha hub; it is a separate component and never sits in the credential path.

No. Ogun contains enough durable orchestration to run acquisition and data workflows by itself, but its job is materializing trustworthy artifacts. It is designed to be called from orchestrators as a task, resource or operator, not to replace them.

As a JSON intermediate representation: a DAG of operator nodes with inputs, configuration and policy. In this release you write the JSON (the VS Code extension and the language server help), or your agent drafts it over MCP. YAML and a Python SDK are planned and will compile to the same IR. Every route ends at the same validator, a plan diff and, for anything sensitive, a human approval.

No. A workflow such as download, unzip, CSV to Parquet runs with no model access at all. LLM operators are optional and run through a provider ladder you configure, with a spend ledger and budgets.

Yes, as a python operator. Your code is referenced by content hash, runs in a sandboxed process with declared inputs and outputs, and adding or changing it in a workflow needs your approval. The sandbox is bubblewrap on Linux; Windows has none yet, so code operators run there only with the explicit allow_unsandboxed override, and the hosted hub cannot run them.

The workflow is hashed; operators are pinned by version and content hash; inputs are content-addressed artifacts; model calls record the model and a prompt hash; and every output records its parents. Re-running a version with the same inputs is a lookup, not a guess.

Completed work is kept. Transfers resume from the engine’s journal, node instances resume from checkpoints, and completed partitions are skipped. That is the phase O1 exit test, which the runtime passes on Linux: a partitioned backfill killed with kill -9 resumes only its incomplete partitions.

Only through Zeus. Ogun never holds browser cookies. A workflow node that needs a logged-in source asks the Zeus daemon, as a service client, to run that acquisition; you approve the site once on a Zeus trust surface, and Ogun imports the result with its provenance.

On your machine, in Ogun’s content-addressed store and in the workspaces you choose. Nothing is uploaded to Verne. Credentials for data sources stay on your machine, zeroized in memory and never logged.

Yes. robots.txt is honoured by default, requests are spaced per host, concurrency is limited, and there is no captcha solving, fingerprint spoofing or DRM removal. If a source needs a person, the run waits for one.

Adapters for Spider 2.0-DBT, ADE-bench and Terminal-Bench 2.0 tasks in the Harbor format are written; their task formats are still being checked against the official releases. Official evaluators run unmodified and their versions are recorded. Ogun does not invent a competing task format.

Each run gets an immutable environment in a container: network denied by default, the workspace read-write, the benchmark read-only, model credentials only, fresh memory and an explicit cache mode. Benchmark runs never see your artifact stores, the Zeus daemon or browser sessions.

Because a run is already workflow plus environment plus agent plus policy plus inputs. Add an evaluator and a benchmark is a controlled set of runs. The same lineage, trajectories and cost ledger then make results reproducible and comparable across agents, models and harness configurations.

Yes, through ogun mcp, which speaks MCP over stdio and offers sixteen tools: validate, diff, partitions and operators without a daemon, and with one, storing workflows, starting runs and reading run, partition, backfill, approval and stream status. An A2A bridge is planned. Agents can propose and run approved workflows; they can never approve anything themselves.

Every LLM operator needs a budget in the workflow. Spend is counted in integer micro-dollars in a hash-chained ledger and enforced per run; raising a budget in a new version needs a human approval.

Verne Ltd., a forward-deployed AI engineering firm in Dhaka, Bangladesh. Ogun is developed in the open under Apache-2.0. See About.

Ògún is the Yoruba orisha associated with iron, metalwork and the making of tools. The name honours that craft: a forge where raw material becomes something reliable. We do not depict the orisha or use sacred symbols; our imagery is public-domain paintings of forges and ironworks, translated into pixels, credited on the art credits page.

Read the architecture and decision records, try the pre-alpha and write to info@vernesystems.com with feedback, and follow Verne on LinkedIn or X. The source is public at github.com/ArefinAlter/ogun. Security reports: see the security policy.