One record from request to production
Frame sets the rules and the plan. Weave agents build the change and open the pull request. Trace keeps the record of every run.
Rules first, then a plan, then code
Each project keeps its rules as control records. Before an agent writes anything it drafts a plan against them, and a person accepts it or rejects it. A rejection must say why and what would fix it, and the next plan has to answer.
- Control records per project, each with how strictly it is enforced
- Rejections carry a category, a reason and a remedy
- ChangeFabric's own rejections wait for a person unless you switch that off
- Ticket
- #418 Retry invoice export
- Category
- Too big
- Reason
- Changes the export and the reconciliation job in one pull request.
- Remedy
- Split it: ship the export retry first.
- Next step
- Re-plan
Agents build it and open the pull request
An accepted plan becomes a change: an agent writes it in an isolated sandbox, runs the checks and opens a pull request in your repository. You choose the model for each automation, and its ceiling bounds what it may spend.
- One sandbox per run, with short-lived credentials
- Model chosen per automation, as a spending decision
- It merges through your own review and branch rules
- Model
- claude-sonnet-5
- Runs in
- An isolated sandbox, one per run
- Ceiling
- $20 and 10 runs per 24h, 2 at once
- Output
- A pull request in your repository
Every run keeps its record
One run, read two ways. Finance sees what each model call cost against the run's budget. Engineering sees the steps, the checks on the pull request and the commit they ran on.
- Each model call with its model, tokens, cost and time
- Governance checks with their violations, per pull request
- Kept in full, with secrets removed, and a one-line summary beside the numbers
- Governance check · 0 violationsPassed
- Unit testsPassed
- Read ticket, draft plan$0.09
- Write the change$0.46
- Run checks, open PR$0.17
What each executive gets back
The same record, read four ways. Each card lists only what the product does today.
CTO / Head of AI
Owns what gets built
AI-built changes arrive faster than anyone can review where they came from.
“What is being built with AI, and who owns it once it matters?”
- Included: Run records
- Included: Rules per project
- Included: Plan review
CFO / Audit
Owns the spend
AI spend shows up as one line on a provider invoice.
“What did each AI change cost, and against which budget?”
- Included: Cost per model call
- Included: Budget per run
- Included: Spend ceilings
COO / Operations
Owns the operation
Automations run the business with no one on the hook for them.
“Which AI work is running, and who answers for it?”
- Included: Ticket timeline
- Included: Roadmap
- Included: Automation limits
CISO / Risk
Owns the exposure
Agents reach code and systems before security knows they exist.
“Which change touched what, under which rule, and who approved it?”
- Included: Sandboxed agents
- Included: Short-lived credentials
- Included: Checks on every PR
GitHub today
We list what works now and say plainly what doesn't.
- GitHubIssues, pull requests and checks, through the ChangeFabric GitHub App.Live
- Other trackers and toolsPlanned. Not available today, and we won't list them until they are.Planned
Security posture, stated plainly
What is in place, and what is not yet. No badges we haven't earned.
- Agents run in an isolated sandboxEach agent run gets its own sandbox, separate from your systems.In place
- Short-lived credentials for agentsAutomated agents get self-expiring credentials, so a leaked one stops working on its own.In place
- Run records keep secrets outWhat agents send, receive and do is kept in full, with secrets removed.In place
- Limits before autonomyAn automation with no ceiling set does not run.In place
- SOC 2Not certified. We don't claim a certification we don't hold.Not yet
Straight answers
Request access
ChangeFabric is in a closed beta with a small group of design partners. Tell us where AI work already runs in your business.
- A person on our team reads every request.
- The first conversation is about the one workflow you name, not a generic demo.
- If it fits, we scope a design-partner engagement and price it with you.