dashcam

The dashcam for your AI coding agents.

A dashcam doesn't stop the car. It records.

The problem

Your coding agents take hundreds of actions a day — reading files, running commands, calling models. When something goes wrong (a bad deploy, a surprise bill, an auditor's question), you have no tape:

• Harness logs exist but nobody reads them. Claude Code writes every session to ~/.claude/projects/, but it's raw JSONL exhaust — no digests, no cost, no verification.
• Your SIEM doesn't see it. Agent activity lives on developer laptops. Nobody pays Splunk ingest pricing for "agent ran ls 400 times" — so the tool-call tape simply never reaches the SOC.
• Cost is invisible until the invoice. Tokens scatter across providers and models with no per-session attribution.

dashcam reads the logs your agents already write and keeps tamper-evident receipts: every tool call, every token, every dollar. Read-only. Local-first. Nothing leaves your machine unless you say so.

What it is

• Session receipts — every agent session parsed from the harness's own logs into local SQLite, each row hash-chained (sha256(prev_hash + record)). dashcam verify proves the tape wasn't altered.
• Multi-harness — Claude Code, Codex CLI, Gemini CLI, and opencode, each with a verified parser. dashcam ingest reads them all; dashcam ingest --harness codex reads one.
• Cross-provider cost — token usage converted to USD per provider/model from a community price table. No proxy, no MITM, no behavior change.
• dashcam today — the daily digest: sessions, tool calls, files touched, tokens, spend — plus unusual-session flags when a session costs or calls tools far above its project's baseline. dashcam today --project foo filters to one project.
• Signed bundles — dashcam export --format bundle --session <id> emits an Ed25519-signed portable receipt anyone can verify offline with dashcam verify-bundle. The artifact you hand to an auditor.
• dashcam serve — a localhost-only read-only timeline UI: today's stats, anomaly flags, recent sessions, per-session detail with chain-integrity check.
• dashcam brief — optional BYOK narration: your own API key (OpenAI-compatible or Anthropic) turns the digest into a plain-English briefing. Off by default; only aggregate stats are ever sent.
• SIEM-ready — dashcam export --format hec emits Splunk HEC events. dashcam is the edge sensor; your SIEM stays the system of record. See docs/SIEM.md.

What it is not

• Not a firewall. dashcam never blocks, prompts, or restrains your agent. Blocking sidecars get bypassed by the agent, fatigue the user, and get uninstalled — we chose the other side of that trade on purpose. The tape doesn't lie, and it doesn't nag.
• Not a cloud service. No accounts, no servers, no telemetry. Receipts live in ~/.dashcam/receipts.db.

Already have Splunk?

Then you know the gap: your SIEM sees EDR process events and proxy logs — that something ran — but not the agent's tool-call sequence. dashcam records the tape where the work happens (free, zero-config) and forwards clean JSONL to HEC when you're ready:

dashcam export --format hec | curl -H "Authorization: Splunk $HEC_TOKEN" \
  --data-binary @- https://your-splunk:8088/services/collector/event

Regulators are asking for exactly this artifact: the EU AI Act (Art. 12, enforceable since Aug 2026) requires high-risk AI systems to keep tamper-evident event logs; SOC 2 and ISO 42001 demand the same discipline. dashcam produces the receipt; your SIEM keeps it. Full mapping in docs/SIEM.md.

Quickstart

pipx install dashcam-ai        # or: pip install dashcam-ai
dashcam demo               # 60-second tour on a scratch DB — your receipts stay clean
dashcam ingest             # read all your agents' session logs (read-only, resumable)
dashcam today              # what did my agents do today? (+ unusual-session flags)
dashcam serve              # localhost timeline UI at http://127.0.0.1:8321
dashcam brief              # BYOK plain-English briefing (needs your API key)
dashcam runs               # sessions with tool calls, tokens and cost each
dashcam verify             # prove the receipts weren't tampered with
dashcam export --format bundle --session <id>  # signed receipt for auditors
dashcam export             # dump receipts as JSON (or --format hec for Splunk)

How it works

flowchart LR
    A["Agent transcripts<br/>Claude Code · Codex · Gemini · opencode"] -->|"read-only parse<br/>(nothing is modified)"| B["dashcam ingest"]
    B -->|"hash-chained,<br/>one ACID txn per file"| C[("SQLite<br/>~/.dashcam/receipts.db")]
    C --> D["today · brief<br/>(digest + anomaly flags)"]
    C --> E["serve<br/>(localhost UI)"]
    C --> F["verify<br/>(chain check)"]
    C --> G["export --format hec"]
    G --> H["Splunk / SIEM"]
    C --> I["export --format bundle"]
    I --> J["Signed receipt<br/>(auditors, clients)"]

Each recorded row seals the previous one into a SHA-256 chain. Delete or alter a row and dashcam verify fails — the same tamper-evidence idea behind signed audit logs, running entirely on your laptop. Details in docs/ARCHITECTURE.md.

Security

Full threat model in SECURITY.md. The short version:

• Tamper-evident, not tamper-proof — stated honestly, with the upgrade path documented.
• Secret redaction — command text is scrubbed for recognizable secrets before persistence; full prompts are never stored.
• Local-first — receipts never leave your machine. The only network calls dashcam ever makes are the ones you explicitly opt into: brief with your own API key. No telemetry, ever. DB is 0600, directory 0700.
• Crash-safe — per-file ingestion is one ACID transaction; concurrent writers can't fork the chain (WAL + BEGIN IMMEDIATE).

Cost data

Prices live in src/dashcam/data/models.json (USD per 1M tokens). Providers change prices; when they do, send a PR that touches only that file. If the table is stale, drop your own ~/.dashcam/pricing.json (same format) — it overrides the packaged table without a reinstall.

Roadmap

• v0.1 — the tape: Claude Code transcripts, hash-chained receipts, cost, digest, demo, verify, SIEM export
• v0.2 — universal tape: Codex CLI + Gemini CLI parsers, unusual-session flags, Ed25519 signed bundles
• v0.3 — opencode parser (SQLite), dashcam serve localhost timeline UI, dashcam today --project
• v0.4 — dashcam brief: optional BYOK narration (your key, off by default, aggregates only)

The roadmap is complete. What's next is driven by real usage — file an issue with what your agents do that dashcam doesn't capture yet. See docs/ROADMAP.md.

Honest comparisons

• vs. blocking sidecars (pre-execution firewalls): they restrain, we record. Restraint gets bypassed by a capable agent and fatigues users; recording stays out of the agent's way and never nags. Different jobs. (Honest scope note: we only see what the harness actually writes to its logs.)
• vs. session scorers (e.g. agent-dashcam): they grade your agent's quality; we keep the evidence. The tape comes before the verdict.
• vs. agentmetry (closest in spirit — local-first flight recorder, hash-chained): they go wider (MITRE mapping, detections, DLP, blocking modes) via IDE hooks; we stay narrow (receipts + cost + digest) and read-only — no hooks to install, none to bypass. Early days for both; pick the philosophy you want.

Contributing

PRs welcome. The bar: deterministic, read-only, tested. Every parser change needs fixture transcripts in tests/; price updates touch only the pricing file. See CONTRIBUTING.md.

License

MIT — see LICENSE.