01
During the run
Record
RunLedger captures every step of an agent run: files read and written, commands, API calls, model and tokens.
Early stage · Building in public
Every agent run, on the record.
RunLedger is a black box recorder for AI agents. It turns each run into a shareable receipt: what changed, what ran, what it cost, and what looked risky, in plain language.
RUN RECEIPT · #rl-0042
Claude Code · repo: payments-service
branch: fix/retry-logic · duration 4m 12s
- 1Read 6 files to understand retry logic
- 2Edited src/payments/retry.ts: added exponential backoff (+38 −12)
- 3Ran npm test — 142 passed
- 4Deleted tests/payments/retry.legacy.spec.ts
- 5Read .env.local
Risk score62 / 100 · Medium
- Deleted a test file
- Touched a secrets file (.env.local)
- No actions outside the working folder ✓
Total 18.7k tokens · $0.071 · Files changed 3
The problem
Agents now ship code. Nobody reads the transcript.
01
Runs happen unsupervised
Agents edit files, run shell commands and call APIs while you're in another tab.
02
Logs aren't explanations
Raw transcripts are thousands of lines. Reviewers need the what and the why, fast.
03
Risk hides in the details
A deleted test, a peek at a secrets file, a write outside the project folder: easy to miss, costly to ship.
How it works
Record, explain, then decide.
02
After the run
Explain & score
Each step is summarised in plain language, and the run gets a risk score with the reasons behind it.
03
Your call
Share, approve, roll back
Send the receipt to a teammate, apply permission rules, and roll the run back if something looks wrong.
Features
Everything a reviewer needs, in one receipt.
Shareable run receipts
One link per run: files changed, commands and API calls, explained in plain language.
Cost & tokens per step
See which model handled each step and what it cost.
Risk score
Flags deleted tests, touched secrets, and actions outside the working folder.
Permission rules
Define what an agent may and may not touch.
Rollback
Undo a run's changes in one step.
Built for Claude Code first
Then MCP-based tools and other agents.
Claude-first
Built Claude-first.
RunLedger uses the right Claude model for each job: Claude Haiku writes fast, low-cost summaries of every step, while Claude Sonnet and Claude Opus handle the harder work: evaluating risk and explaining diffs.
- Haiku
- Step summaries
- Sonnet
- Risk evaluation
- Opus
- Deep diff explanations
First integration: Claude Code Next: MCP other agents
Claude, Claude Code, Haiku, Sonnet and Opus are products of Anthropic. RunLedger is an independent project, not affiliated with or endorsed by Anthropic.
Get early access.
We're opening RunLedger to a small group of developers first. Tell us what agents you run and we'll reach out.
Where we are
RunLedger is an early-stage, bootstrapped project founded in October 2026. The product is in active development; the receipt shown on this page is an illustrative mockup, not live data.