agentcamRECA dashcam for your coding agents.
agentcam reads the session logs Claude Code and Codex already write on your machine and shows you what your agents did: every session replayed step by step, searchable, with the code that lasted and the steps worth a second look.
- Claude Code
- Codex
- Read only
It reads the logs your agents already write. It never edits, moves or uploads them.
- Local only
One server on 127.0.0.1. No account, no telemetry, no network calls.
- Nothing to set up
No hooks or SDK. Your past sessions show up the first time you run it.
- One binary
Written in Rust, with the web UI built in. Open source under Apache 2.0.
Replay
Scrub through any session.
A timeline you can zoom and play at 1×, 4× or 16×. The transcript follows along, split into turns, with each command and its output, each diff, how each call was approved and what it cost.
Every step has a link
Send someone the exact edit you want them to look at. Open any changed file at its line in VS Code, Cursor or Zed.
Pick up where it stopped
Copy the command that resumes the session in its own folder: claude --resume or codex resume.
Made for the keyboard
- Spaceplay
- ← →step
- ⇧Jnext file change
- Enext error
- Rnext risk
Did it work?
See what the work came to.
Steps and tokens tell you how busy an agent was. These tell you what stayed. All of it works on your past sessions, with nothing to install first.
Of the lines your agents wrote, how many are still in the code today. By model, project, file and week.
Safety
A flight recorder, not a cop.
agentcam assumes your agent is capable and acting in good faith. It shows what happened and flags only what deserves a second look, in plain words.
Instructions that came from outside.
A fake Sentry error tells the agent to run a package. agentcam notices when a command runs a package or sends data to an address that showed up in a tool result and never in your prompt, and links the two.
It also reads what each MCP server and skill told your agent, and points out hidden characters, orders like “don’t tell the user”, and descriptions that changed after you started using them.
- 73 built-in rules
Secrets in transcripts (masked before they are stored), sensitive files, risky commands, data sent out, writes outside the project. Add your own in YAML.
- Chains across steps
A secret read and then an upload. A stopped call, then another way around it. Each finding links back to where the chain started.
- Approvals
How each tool call was allowed or stopped, and allow rules you could add so agents ask you less. You copy them; agentcam never edits settings.
- Block next time
Turn a step into a guard rule. A PreToolUse hook stops matching calls with a message to the agent. Only rules you set ever stop anything.
Day to day
Keep an eye on the agents you run.
Which agents are working now, which are waiting for you, and how close you are to your usage limits.
- Cost
Estimated per step, turn, session and day, from a price table you can edit.
- Usage limits
Codex's 5-hour and weekly windows, and an estimate of Claude Code's from logged tokens.
- Alerts
A browser notification when a session waits for you or passes a cost you set.
- Ask your history
agentcam mcp lets your agents look up past work, with secrets masked.
- Evidence bundles
A session or date range as one JSON file with a SHA-256 checksum.
- SIEM and OpenTelemetry
Findings as NDJSON or OCSF, and sessions as traces for Jaeger or Tempo.
- Export
Any session as Markdown or JSON, any table as CSV, secrets masked.
- Commits and PRs
The commits each session made and the pull requests it opened.
For teams
A team server you run yourself.
Each laptop sends metadata only: counts, times, cost, findings and file paths. Never prompts, output or diffs. It covers Claude Code and Codex together, on any plan or API key. Names stay hidden unless an admin turns them on, and no view ranks people.
Agent time, cost, lines kept and rules followed, by week, team, repo and model.
Privacy
Your logs stay on your machine.
- The server listens on 127.0.0.1 and turns away requests for any other host name.
- Secrets the rules find are masked before they reach the index, and again in exports.
- It reads project files to count lines that lasted, and stores counts, never contents.
- Nothing leaves until you join a team, and the Outbox page shows exactly what would.
Install
Try it on your own logs.
agentcam is in heavy daily use and not packaged yet. Build it from source with Rust and Node 20+. It indexes your logs, starts on 127.0.0.1:4747 and opens your browser. Packages for cargo, npx and Homebrew are coming.
git clone https://github.com/chauhanswapnil/agentcamcd agentcamnpm --prefix web ci && npm --prefix web run buildcargo install --path crates/agentcamagentcamSee what your agents did.
Open source, Apache 2.0. Issues and pull requests welcome.
chauhanswapnil/agentcam