Witness
Voice-anchored causal evidence recorder that captures spoken observations alongside browser events and hands an evidence package to AI coding agents.
At a Glance
MIT-licensed, local-first CLI recorder, renderer and MCP server; no accounts or required keys.
Engagement
Available On
Alternatives
Listed Nov 2026
About Witness
Witness is an open-source, MIT-licensed CLI (wit) from zalize that records a browsing session while you talk, then ties each spoken phrase to what the page did at that moment. It produces a local evidence package for humans and AI coding agents, and can serve that evidence over MCP.
What It Is
Witness is a voice-anchored evidence recorder for debugging with AI coding agents. Running wit start <url> opens a managed browser (system Edge or Chrome preferred, bundled Chromium as fallback). You use the product and narrate, for example "this button does nothing". Witness attaches clicks, network requests and responses, console errors, DOM mutations and screenshots from that second to your words. A click with no aftermath is recorded as evidence too.
How the Recording Works
Three streams run in parallel: continuous audio with live streaming transcription, operations with roughly 1.8-second causal bursts after each click, and page evolution via rrweb full-DOM replay, screenshots and periodic DOM snapshots. In-browser gestures include Ctrl+click to mark an element, M for inspect mode, N for an issue boundary, P to pause (freezing evidence and mic) and Alt+Shift+R to stop. Events stream to disk so wit recover can rebuild artifacts after a crash.
The Evidence Package
Sessions land in .witness/sessions/. Humans get review.html with speech cards, audio seek, a film strip and embedded DOM replay. Agents get PROMPT.md, SEGMENTS.md, session.md, per-segment bundles, events.jsonl, DOM snapshots, response bodies and screenshots. Speech is kept verbatim; correction and association by an LLM are optional display-layer additions.
Agent Handoff and MCP
wit mcp serves a session over stdio with 12 tools, including list_says, events_around, get_dom_state, get_errors, get_shot, get_body, session_md and get_asr/set_asr. Agents can alternatively read the directory directly or be given PROMPT.md. A SKILL.md ships in the repo for skill-aware agents. The site lists Devin, Claude Code, Cursor, Codex and Windsurf as compatible agents.
Privacy and Local-First Design
Evidence stays on disk and .witness/ is gitignored automatically. Password fields are masked across four channels, POST bodies redact credential fields, and WIT_NO_BODIES=1 disables response capture. Live captions send mic audio to the chosen ASR engine, and the CLI sends anonymous usage pings that can be disabled with WIT_TELEMETRY=0. No API keys are required; OpenAI and Zhipu keys enable optional upgrades.
Community Discussions
Be the first to start a conversation about Witness
Share your experience with Witness, ask questions, or help others learn from your insights.
Pricing
Open Source (MIT)
MIT-licensed, local-first CLI recorder, renderer and MCP server; no accounts or required keys.
- wit CLI: recorder, renderer and MCP server in one
- 12 MCP tools
- rrweb full-DOM replay, screenshots and causal bursts
- Evidence stored locally in .witness/sessions
- Free ASR engines by default; paid-model hooks are optional bring-your-own-key
Capabilities
Key Features
- Voice-anchored evidence timeline with verbatim speech
- Causal bursts capturing requests, responses, console errors and DOM changes after each click
- Dead-click detection recorded as evidence
- rrweb full-DOM replay, screenshots and DOM snapshots
- Element marking with selector, component chain and framed screenshot
- Inspect mode with region selection
- Issue boundary markers and pause/resume
- Streaming ASR with auto-fallback (doubao-web, doubao-ime, browser SR)
- Human review.html with audio seek and film strip
- Agent handoff files: PROMPT.md, SEGMENTS.md, session.md, events.jsonl
- MCP server with 12 evidence-query tools
- Crash recovery via wit recover
- Password masking and body-capture redaction
- Bundled SKILL.md for agents
