Six layers between you and the ledger — and you stay in control of every one of them. AI proposes. Humans (and risk-based rules) merge. If you've read a Stripe API reference, this will look familiar.
AI proposes. Humans (and risk-based rules) merge. Nothing reaches the ledger without passing through GitFinance — the control layer where every write is reviewed, signed, and audited.
Every write is risk-scored at submission. The lane decides whether the ledger gets it now, after one approval, or after a signed multi-party review. The agent never decides — the rules and the humans do.
Low-risk, recognized pattern, trusted operator. Posts immediately, logs the decision, no human in the path.
Moderate amounts, edits to master data, new counterparties. Routes to a single reviewer with full context attached.
High-risk or irreversible. Two or more reviewers, statutory sign-off where required. The agent never bypasses.
Agents hand work to each other directly — no central queue. When one agent hits a missing prerequisite (an unknown vendor, a closed period), it asks the agent that owns that domain to fix it, then resumes via callback. Human gates appear in the same trace as H rows.
Each plugin is a directory under mcp-server/agent-definitions/ or a top-level plugin folder. MIT-licensed. Forkable per firm, per client. The runtime executes the workflow, GitFinance gates the writes, the ledger receives only what was approved.
One ledger, one control layer, many places to interact with them.
User-facing capability, counted from the running codebase — not a roadmap.