Riven Forge
BetaThe build and code-execution service behind agentic work — sandboxed workspaces, repository operations, and reproducible runs.
Forge is where code actually runs. When an agent in Riven Computer clones a repository, edits files, runs a test suite, or produces a build, Forge provides the isolated workspace and executes the work.
Workspace model#
- A workspace is an ephemeral filesystem plus a pinned toolchain, created per task and destroyed when the task ends.
- Network egress is deny-by-default; a workspace reaches only the endpoints its task declares.
- Repository access uses short-lived credentials resolved from Riven Keymaster — no long-lived tokens land in the workspace.
- Filesystem and command tools are exposed to the model through Riven Cortex, not through direct shell access to a shared host.
Reproducibility#
| Pinned | Why |
|---|---|
| Base image digest | The same task run twice starts from identical bytes |
| Toolchain versions | A dependency resolution change cannot silently alter a build |
| Task definition | The command sequence is recorded with the run |
| Model and routing decision | Recorded via x-riven-upstream so a result can be attributed to a pool |
Forge is in Beta. Long-running and scheduled builds are supported, but capacity is not yet reserved per organisation — a burst can queue behind other work. Interactive agent runs are prioritised over batch builds.
Relationship to other services#
- Riven Exec schedules and supervises Forge runs.
- Riven Fleet Scheduler allocates compute when a run needs GPU access.
- Riven Relay publishes the artifacts a run produces.
- Riven Telemetry carries run logs, durations, and failure traces.