ResearchForge for teams
When experimentation becomes maintenance.
Keeping ML systems healthy eats your team's time: dependency upgrades, retraining on new data, swapping models and prompts, chasing regressions. ResearchForge runs that work as controlled jobs against a frozen baseline — and only what passes reaches your branch.
Where the cost hides
None of this is research. All of it needs the same discipline.
A fair baseline, an isolated test, a clear pass or fail, and someone signing off.
Today
- A dependency upgrade quietly changes model output and nobody notices for weeks
- Retraining on new data is a manual, half-remembered process
- Trying a cheaper model means a week of ad-hoc comparisons
- "It got better" with no fixed baseline to prove it
- Nobody can say what last quarter’s fixes cost in compute
With ResearchForge
- Each task runs against the same frozen baseline, in an isolated worktree
- Hard constraints fail a change automatically — accuracy, latency, cost
- Every attempt, rejection and reason stays on record
- A human approves before anything ships, as a clean branch with a report
- Compute and token cost are measured per project, not estimated
Status, plainly
What exists today — and what doesn't yet.
Every capability carries its real status. “Pilot” means we build it with your team.
Frozen baseline
Measured before any change runs — optionally the mean of N runs, with the spread recorded.
Isolated runs
Every attempt runs in its own git worktree (venv or Docker) from the same baseline commit. Your branch is never touched.
Protected paths and hard constraints
Changes to protected files are rejected; latency, cost or accuracy limits fail a result automatically.
Human approval gates
Contracts and plans need a typed approval; opening a PR on a repo you cannot push to always asks.
Audit trail
audit log / audit export rebuild the history from the records the workflow writes, and flag plans that ran without an approval.
Your choice of model
Anthropic, Gemini, OpenAI, or a local model through Ollama — or drive it from Claude Code / Cursor.
Measured compute and token cost
Compute by stage and by outcome, token spend per project and purpose. Unknown prices are named, never shown as zero.
Scheduled background runs and alerts
A monitor that runs on a schedule and pushes results to your phone; a live local dashboard.
Recurring maintenance jobs on your repo
Dependency upgrades, retraining on new data, model or prompt swaps — each run as a job against the frozen baseline.
Shared history across a team
One record of who ran, approved and shipped what.
SSO, CI/CD plugin, tracker integrations
Built when a pilot needs them.
30-day pilot
Bring us one repository and one recurring task.
We set up the baseline, constraints and approval flow on your repository and run your maintenance task through it. You keep the full record and the measured cost of every run.
Frequently asked
Is the open-source version really free?+
Yes. The CLI and the Claude Code / Cursor skills are Apache 2.0, free for individuals and companies, including commercial use.
What does a pilot look like?+
You bring one repository, one metric and one recurring maintenance task. We set up the baseline, constraints and approval flow with you and run it for 30 days. At the end you have the record of every run and its measured cost, and you decide whether to continue.
Does our code leave our infrastructure?+
ResearchForge runs where you run it and sends nothing to us. The only outside calls are the ones you configure: your AI provider (or a local model), arXiv for literature search, and notifications if you turn them on.
Do you have SOC 2 or other certifications?+
No. ResearchForge keeps a complete record of every experiment, decision and approval, which can support your own evidence collection — but we do not hold a certification.