Specialist agent foundry
Forge digital employees that earn.
Select the right base model, train a private specialist, deploy it as an endpoint, then publish the agent into a metered marketplace.
The forge method
A premium production line for owned agent labor.
Select
Rank base models by quality, cost, license, and task fit before committing spend.
Prepare
Turn examples and documents into trainable rows with quality gates and privacy checks.
Train
Launch adapter jobs, watch loss curves, and compare tuned behavior against the base model.
Sell
Deploy an endpoint, publish an agent card, and meter calls into a revenue stream.
Adapter to revenue
Turn a workflow into a productized employee.
Forge is where a member’s best examples become reusable behavior. Every stage is designed to answer the buyer’s question: does this agent perform, can I trust it, and what does it cost?
Quality gatesDeduplication, leakage, PII, evals, and base-vs-tuned evidence.
Deployable endpointsPrivate API keys, usage limits, MCP surfaces, and metering.
Marketplace economicsRental pricing, adapter rights, seller rewards, and trial access.
Define the employee
Describe a valuable narrow job and the examples that prove quality.
Forge the adapter
Choose the base model, pass the data gate, and train a specialist behavior layer.
Meter the work
Deploy private endpoints or publish usage-priced agents buyers can discover.
Live forge studio
Build, train, deploy, and list from the same surface.
0
specialists deployed
1626
synced corpus docs
0
FractalChain credits / call
0
router eval datapoints
Pillar A
Model Selector
Bake-off queued for 3 examples. The current scorer is deterministic until inference workers are connected.
Ranked shortlist
Candidate base models
Run the selector to generate a ranked shortlist.
Pillar F
DataMesh
Google DriveReady1240 docs
NotionSyncing386 docs
GitHubPlanned0 docs
SlackPlanned0 docs
External runtimes
Hermes / OpenClaw agents
- Connect Hermes or OpenClaw runtime
- Import normalized harness and approve permissions
- Pass eval gate before routing or marketplace listing
Pillar B
Fine-tune engine
- Data quality gate: dedup, leakage, PII report
- LoRA/QLoRA training job with live loss curves
- Base-vs-tuned eval card before deployment
Pillar C
Deploy & meter
- Promote adapter to private endpoint
- Issue per-agent API keys and limits
- Expose the agent as a remote MCP tool
Pillar D
Marketplace
- Agent card required before publish
- License chain blocks invalid terms
- FractalChain rental rewards settle through batch roots
Router index
Indexed agent documents
PRD implementation tracker
Phased roadmap
| Phase | Scope | Status |
|---|---|---|
| Phase 0 | Selector, JSONL/CSV training, private deploys, MCP call wrapper | Now |
| Phase 1 | Bake-off runner, raw-document synthesis, first marketplace rentals, Drive + Notion sync | Next |
| Phase 2 | License templates, payouts, eval-card v2, GitHub + Slack, portable memory | Planned |
| Phase 3 | Learned /route endpoint and FractalWork graph assignment integration | Later |
Ship the employee