Forge (control plane)
The browser app: workspace tabs, Pipeline, agent fleet, Mission Control, approvals, health and history. Owns the human experience. Stops at executing tools — it only reads/writes behavior through validated agent JSON.
Product overview
One platform,
clear boundaries.
Forge is a small number of systems with sharp responsibilities. Every box below states what it owns and where it stops — so behavior stays predictable as the platform grows.
7workflow verticals
5runtimes
11tools and integrations
4permission tiers
A
Surfaces
Forge · Studio · Weave
Forge Studio
The visual agent composer: drag-drop canvas, node inspector, lint, simulate, version and publish. Owns authoring. Stops at execution — it converts agent JSON ↔ graph losslessly and hands off to the runtime.
Weave
Context- and screen-aware assistant, grounded by RAG, answering in natural language. Owns explanation and navigation. Stops at mutation — Weave is strictly read-only.
🐕
Live companion
Forge Pet
Idle
Meet the Forge dog — a tiny companion in the corner of every screen (that's the real sprite animation on the left). At a glance it shows what the platform is doing: it perks up when work is running, celebrates a shipped result, and dozes when things are quiet.
Click it to jump straight to whatever needs you — an approval, the running fix, the pipeline or the queue. It's playful, never in the way, and can be hidden or compacted. Best of all it's strictly observer-only: it just watches what the runtime is already doing, so it can never touch a run.
B
Engine
Fabric — decision, knowledge, execution
Decision Engine
Pure, auditable policy: agent/runtime routing, confidence, preconditions, roles, retry and plan. Boundary: confidence can request review but never overrides a hard failure.
Knowledge Engine (semantic + structural)
Semantic: workspace-scoped RAG over runs, KB, agents, skills and registered sources, plus scoped memory — local (JSONL) or hybrid (pgvector) mode (ADR-0001). Structural: a Code Intelligence Engine (ADR-0003) answers what does changing this touch — impacted symbols, call chain, blast radius — from a code graph, behind a provider seam (Graphify is the first provider, swappable). Semantic tells Fabric what the code means; structural tells it what the code touches. Evidence carries a trust tier — trusted (our deterministic bytes/git), internal (vetted systems like Jira/Bugsnag), derived (advisory findings computed from trusted bytes, e.g. the code graph), untrusted (free-form text, quarantined). Boundary: derived and untrusted content grounds but never authorises; missing evidence stays missing.
Execution Runtime
Five runtimes on one engine — generic, unity, offer, data and engineering (the governed bug-fix engine) — dispatching stages to registered executors and streaming progress. Boundary: a capability runs only where an executor supports its mode and runtime.
AI capability router
Maps a capability (reason · code · retrieve · validate · vision · embed) to a provider/model, deterministic-first. Boundary: models are replaceable plugins — swap them and nothing above moves.
C
Definitions
Agents & pipelines
Agents
A JSON definition: trigger, inputs, context, knowledge, stages, routing, tools, policies, validators, confidence thresholds, approvals, outputs and provenance. The single source of truth between authoring and execution.
Pipelines
The staged flow an agent runs — Observe → Reason → Plan → Execute → Learn — with validation, human approval and evidence capture before any outward effect. See a real Jira→output walkthrough in Pipelines & runs.
Seven workspaces; live agents in one.
Seven verticals are stood up — Tech Art, Design, Production, QA, Product, Dev and Art — with live agents today in one: Tech Art (3 live in production, with more of the defined roster — Ability, Flat Art Import, New Flat Art Setup, Offers-DD, DD Iconic Item, PG Setup — rolling out). Design (Design Character Setup) and the Dev vertical's governed engineering / bug-fix engine (incident monitoring + autonomous fix) are in development, not yet live. The remaining workspaces have the full shell and commission domain agents through Studio as each contract is ready — every new workflow becomes a new agent on the same runtime, not a new tool.
D
Trust
Governance, observability & dashboards
Harness & approvals
Read-only validation, simulation and certification; role-gated human sign-off before outward effects. Includes the Character definition-of-done readiness check.
Observability & audit
Live NDJSON run streams, telemetry, health, run journal and failure post-mortems. Every deploy/admin action is audited.
Production dashboards
Mission Control, Queue, Completed, Hanging Work and History give a live and historical view of every execution. See the screen guide.
E
Edges
Integrations, security & deployment
Integrations
Jira/Confluence, Perforce, Adobe Photoshop, Unity, Git/GHE/S3, Google Sheets/Apps Script, Slack, Bugsnag and SmartBear MCP — each connector keeps its native protocol behind a shared resilience layer.
Security & deployment
Origin/Host/CSRF guards, per-user tokens and roles (viewer → operator → deployer → master), audit trail, private transport. One FORGE_PROFILE selects the shape: local (one private workstation) or hybrid (shared control plane + authenticated workers), both shipped as a repo-free signed package (ADR-0002). (SSO: Planned.)