Observe
See the plan, active cycle, tool activity, and result as it happens.
Designed as a system
Every mode of work shares the same context and controls.
Chat, goal loops, and automation use the same context engine, approvals, skills, apps, and durable state.
- One local gateway
- One web dashboard
- One inspectable home directory
The 84-second tour
One unedited run through the parts that matter.
A single recording of v0.1.3 on one machine — the version in the capture's status bar is the release this site publishes. Chat against a local model, the approval it stops on, a goal loop, the local corpus, and the file the agent actually wrote. No narration, no cuts, nothing sped up.
- 01 Chat against a model on this machine
- 02 The approval it stops on before writing
- 03 A goal loop with its findings
- 04 Knowledge searched locally
- 05 The artifact it wrote, on disk
Agentic chat
Conversation that can become work.
Run tools with approval controls, organize sessions, branch from earlier turns, compare answer variants, and tune model or memory behavior per conversation.
- Session fork and undo
- Folders, tags, and kanban
- Temporary and incognito memory modes
- Per-session model overrides
Autonomous goals
The agent keeps going. You keep the final say.
Intervene
Pause, nudge, resume, or stop without losing the accumulated run state.
Bound
Deterministic supervisors and approval modes constrain autonomous action.
Durable context
Memory and knowledge do different jobs.
PersonalClaw keeps learned personal context separate from the source material you deliberately add. Both remain inspectable.
Memory learns how you work.
Semantic, episodic, and procedural layers support recall, correction learning, and promotion of repeated facts.
Knowledge keeps the source.
Ingest documents, web pages, and media; enrich them with entities and make them searchable from chat.
Automation
Let the system notice first.
Schedule recurring work, react to webhooks, monitor connected channels, and dispatch background subagents without building a second automation stack. Anything that needs you lands in one inbox you can answer in place — a loop waiting on a decision, a proposed skill, an approval you walked away from.
- Cron, interval, and webhook triggers
- One inbox for decisions, approvals, and messages
- Delivery per kind: notify, badge, digest, or never
- Background subagents
- Automatically matched workflow SOPs
One place to answer
Work that waits on you, in one queue.
A loop needing a decision, a proposed skill, an approval you walked away from — these used to be toasts that scrolled past while the work stayed stalled. Now they are items you answer in place.
Everything waiting, together.
Decisions, proposals, approvals, and messages share one inbox. Each row deep-links to the loop or chat it came from.
You choose what interrupts.
Per kind of notification: notify, badge, batch into a daily digest, or never. Quietening one noisy kind no longer raises the bar for everything.
What it produced
Output you can find again.
Documents, spreadsheets, charts, and diagrams the agent made stay in a library with live previews — searchable, versioned, and grouped into collections instead of scattered through a chat log.
- Live previews, not placeholder tiles
- Version history and diffs between revisions
- Collections, and dedup at save time
Bring the right agent
Native, ACP, or app-defined.
PersonalClaw can run its native agent or connect external coding agents through ACP. Claude Code, OpenAI Codex, and Kiro CLI remain behind the same host approval boundary.
Explore agent apps
Multiple runtimes ACP boundary Shared context Replaceable providers
Runs safely on its own
Autonomy with a floor under it.
Before you hand the system unattended work, it gives you the controls to trust it: a kill switch that suspends every automation at once, daily spend ceilings, outbound secret scanning, and per-provider circuit breakers. Interactive chat is never affected.
- Incident kill switch, explicit resume
- Token and dollar budgets per day and per run
- Outbound prompt scanning: warn, redact, or block
- Named safety profiles for unattended runs
Diagnoses itself
See what's healthy, and fix what isn't.
The Doctor runs read-only health probes across every subsystem — memory, channels, local models, app backends, the SPA, model-provider breakers. A degraded capability is never mistaken for a dead gateway, confirm-gated fixes repair the common faults, and a health-scored engine keeps the stores tidy on its own.
- Tiered probes with per-capability evidence
- No-model degraded mode, stated plainly
- Confirm-gated repairs and crash capture
- Health-scored self-maintenance
Open source · self-hosted
Run the system on your machine.
Python 3.12, an aiohttp gateway, React dashboard, SQLite state, and no product telemetry.