GoClaw can now connect to external MCP servers and use their tools as if they were built in — configure a server, and its tools show up in your role permissions automatically. Cron jobs that should stay silent are also more reliable now, even when models get chatty.
New Features
MCP Client Support
Connect GoClaw to any MCP server — remote over HTTP or local via subprocess — and their tools appear as regular GoClaw tools. Access is controlled through the existing role system, and you can filter which tools to import per server with allow/deny lists. Manage connections live with /mcp or the HTTP dashboard.
Silent responses from cron jobs now work reliably even when the model appends extra text before SILENT_OK. Previously this could leak internal narration into your chat — now the entire response is suppressed as intended.
v0.1.20
Background Delegation, Account Login & /btw
GoClaw's agent can now hand off coding work to Cursor or Grok in the background and keep chatting with you while it runs — you can ask what it's doing at any time. You can also log in with your xAI or OpenAI account instead of pasting API keys, and ask quick side questions to a working agent without derailing it.
New Features
Background ACP Delegation
GoClaw's agent can start a coding task on an attached Cursor or Grok session and stay responsive while it works. Live progress streams into your chat, and GoClaw reports status when you ask. When the coding agent finishes, GoClaw picks up automatically.
Ask the attached ACP agent a quick question without breaking its flow. /btw interrupts, answers with full context, then the agent resumes what it was doing.
Sign in with your xAI or OpenAI account instead of managing API keys. GoClaw stores tokens encrypted and the xAI image, video, and search tools reuse the same login. Set up via device code (great for headless servers) or browser OAuth.
The config editor now runs the entire login flow inline — pick auth method, start login, paste code, done. No need to leave the editor or juggle browser tabs separately.
Immediate Cancellation
Cancelled or timed-out agent runs now stop right away instead of being misread as slow-provider timeouts. Deadlines are honored promptly and stuck runs can't keep burning models past their budget.
Cron Stale-Run Auto-Reap
Jobs that crash mid-run no longer appear stuck 'running' forever. GoClaw clears leaked running flags automatically after a configurable ceiling (default 30 minutes).
New allowPrivateNetworks option lets the browser reach LAN devices (like a local Home Assistant) when enabled. Loopback and cloud-metadata addresses stay blocked.
GoClaw can now drive Grok Build sessions alongside Cursor — attach to whichever agent fits the task. The installer and self-updater have moved to goclaw.org so updates keep flowing without GitHub. Error handling for xAI models is also smarter: fewer false alarms, better failover.
New Features
Grok ACP Driver
Attach to xAI's Grok Build CLI as an ACP session, just like Cursor. Pick your driver per attach or set a default. Grok is now the default driver — switch back with /acp attach cursor or change acp.defaultDriver in config.
While an ACP session is attached, the HTTP chat header shows live token usage from the agent — including context window percentage, active model, and session state.
Slash commands the attached agent advertises (like /compact or /help) now go to the agent instead of GoClaw. You always keep /acp, /a2a, and /shutdown locally.
The install script and self-updater now pull releases from git.goclaw.org. Existing 0.1.18 installs self-update automatically; older installs need a one-time reinstall via curl -fsSL https://goclaw.org/install.sh | sh.
Models that don't support reasoning effort no longer trigger a misleading 'session corrupted' error — GoClaw retries without it automatically. Backend errors also fail over to the next model instead of killing the conversation.
Grok Status Notices
When Grok retries a failing API call, a subtle status notice appears in HTTP, TUI, and Telegram so you know it's working on it. Rich error messages replace the generic JSON-RPC codes.
Routine libp2p rendezvous and address-list chatter stays out of INFO logs. Real connectivity changes still surface.
v0.1.17
Send Anything from Telegram and WhatsApp
Videos, documents, audio, stickers — Telegram and WhatsApp now accept them all. No more silent drops. Plus better PDF extraction with smarter OCR and cleaner output.
New Features
Full Upload Support
Send documents, videos, audio files, stickers, and animations from Telegram or WhatsApp. They're saved permanently and the agent can work with them.
Send GoClaw a PDF, Word doc, or e-book and it reads the whole thing for you. Routines now fire on schedule and can nudge you at the right time across all your channels.
New Features
Read Any Document
Upload PDFs, Word docs, spreadsheets, or e-books and GoClaw extracts the text automatically. Scanned pages get OCR. No more copy-pasting.
New docs covering supported formats, OCR, caching, and vision chain behavior.
Memory Graph Updates
Expanded docs for recurring routines, trigger inspection, and the Today's Schedule bulletin.
v0.1.15
Lossless Context Management
GoClaw now remembers long conversations. When the chat outgrows the context window, older parts are rolled up into searchable summaries instead of being dropped. The agent can drill back into the original messages whenever it needs the detail.
New Features
Lossless Context Management (LCM)
Compacted history is now stored as searchable summaries linked to the raw messages they covered. Use grep_summaries, describe, and expand to recover details from months ago.
Four presets — balanced, aggressive, long_term_memory, and recall_heavy — control how much history GoClaw carries forward. Configure under session.summarization.compaction.lcm.preset.
Comprehensive docs covering LCM architecture, presets, condensation, recall tools, and configuration reference.
Transcript Search Guide
Updated transcript search docs with grep_summaries, describe, and expand action examples for compacted history recall.
v0.1.14
A2A Networking, Local LLMs, and Expanded Cursor ACP
GoClaw can now talk to other GoClaw nodes over the internet and run models locally without external providers. A2A networking uses libp2p with relay fallback and NAT traversal. Local LLM management downloads, configures, and runs llama.cpp models with prompt caching.
New Features
A2A Peer Networking
Send tasks to remote GoClaw nodes over libp2p. Includes rendezvous discovery, relay-first connectivity with automatic direct upgrades, and NAT traversal. Use /a2a commands or let the agent handle it.
Run models locally with managed llama.cpp. GoClaw handles model downloads, server lifecycle, prompt caching, and hardware detection. Setup editors guide you through model selection.
New dedicated docs covering A2A setup, peer pairing workflow, configuration reference, troubleshooting, and agent tool usage.
v0.1.12
ACP Configuration Section, Setup Editors, and Dashboard Shortcuts
GoClaw 0.1.12 adds a dedicated ACP config section with setup editor support, live Cursor model refresh, and new dashboard shortcuts for Coding Agents and Web Search.
New Features
Dedicated ACP Configuration Section
Cursor ACP preferences now live under a top-level `acp` config section with browser and TUI setup editor support.
New operator and tool reference docs for Cursor attachment, session scoping, and agent ACP workflows.
v0.1.10
Script-friendly status, restart helper, and clearer installer guidance
GoClaw 0.1.10 makes automation around installs and operations easier: structured status output for scripts, a small restart command for the daemon, and installer messaging that better matches already-configured and already-running setups.
Improvements
Script-Friendly Status and Restart
goclaw status --field provides shell-safe, structured output including a clear configured=false path when no config exists yet; goclaw restart offers a simple daemon restart helper.
0.1.9 TUI Onboarding Flow Alignment and Setup Validation
This 0.1.9 follow-up aligns TUI onboarding flow with the browser wizard, improves step progression and validation behavior, and tightens setup requirements around channel, provider, and voice configuration.
Improvements
TUI Onboarding Step Flow Aligned with Browser Wizard
TUI onboarding now follows the same communication-channel, pairing, provider, and review progression as the browser setup experience for a more consistent operator workflow.
Dynamic Step Management for Conditional Setup Paths
Wizard step handling now supports conditional step insertion/removal while preserving user position, improving behavior when setup choices change mid-flow.
Setup now requires explicit preset warning acknowledgement before advancing, preventing accidental progression past important sandbox warnings.
Required Setup Field Validation Before Step Exit
Wizard exit checks now enforce required fields for enabled channels and selected providers, reducing incomplete setup saves.
v0.1.9
Channel Pairing in Setup, Installer Guidance, and 0.1.9 Security Updates
GoClaw 0.1.9 adds first-class Telegram and WhatsApp owner pairing in setup flows, improves setup UX and install guidance for new versus existing installs, and includes a security dependency bump.
New Features
Built-In Telegram and WhatsApp Owner Pairing
Setup flows now support direct owner pairing for Telegram and WhatsApp across browser and TUI onboarding/edit modes, with staged identity data saved when setup is completed.
The installer now detects existing binaries/configs and suggests the right next step (`goclaw onboard` for guided setup or `goclaw setup edit` for existing configs).
Upstream dependencies now include `golang.org/x/image` v0.38.0 to address the reachable TIFF decoding vulnerability identified during audit checks.
Documentation
Configuration and Runtime Behavior Clarifications
Documentation updates clarify setup pairing behavior, installation flow, and additional web chat/runtime details in the latest release.
v0.1.8
Media Tooling, Voice Effects Controls, and Structured Memory Scheduling
GoClaw 0.1.8 adds an agent-facing media inspection tool, expands voice effects configuration in the web setup flow, and upgrades memory graph workflows with structured scheduling and stricter real-time memory-formation guidance.
New Features
New `media` Tool for Storage Visibility
Agents can now inspect live media store usage, quotas, retention policies, and category warnings before generating, downloading, or storing large files.
Memory graph now supports structured future scheduling fields and upcoming-event handling in query and bulletin flows for deadlines, appointments, and plans.
Agent memory guidance now explicitly requires completing recall/store decision workflows before final user responses when memory-worthy information is detected.
Docs were updated across media, memory, configuration, and voice pages to reflect the 0.1.8 runtime changes and operational defaults.
v0.1.7
Browser Remote Profiles, Capture Controls, and 0.1.7 Runtime Refinements
GoClaw 0.1.7 expands browser automation with remote CDP profiles, capture/tracing/emulation actions, and MCP-style aliases, while improving memory extraction coordination and delegated-run reliability signals.
New Features
Named Remote CDP Browser Profiles
Browser automation now supports named remote CDP profiles, host allowlists, and optional HTTP discovery for connecting to trusted Chrome instances on other machines.
The browser tool now exposes console/network capture inspection, performance tracing and metrics, plus page-level viewport/device/CPU/network emulation controls.
Memory extraction now coordinates agent-driven and background extraction with a configurable handoff delay to reduce overlap and improve ingestion timing.
A regression around the `phase` parameter handling in the OpenAI-next driver path was fixed.
Documentation
Browser Documentation Refresh
Browser docs were updated to cover remote profiles, capture/tracing, emulation, and alias behavior for the expanded browser tooling.
v0.1.6
Subagent Documentation Refined for Clarity
This docs refresh simplifies subagent and delegated-run terminology across concepts, tools, roles, and configuration so operators can understand behavior and controls more quickly.
Improvements
Clearer Subagent vs Delegated Terminology
Core docs now use simpler subagent-first language while keeping delegated-run internals in context, reducing ambiguity in everyday usage.
Configuration reference now explains `tools.subagent` and `gateway.delegatedRuns` dependencies and run-limit settings in more practical, user-facing wording.
Tool reference for `subagent_spawn`, `subagent_fanout`, `subagent_status`, and `subagent_cancel` now focuses on operational behavior instead of deep implementation details.
Role docs now more directly explain owner-only subagent access and why role allowlists alone cannot grant delegated-run control to non-owner users.
v0.1.6
Delegated Subagents, Fanout Runs, and Runner Visibility
GoClaw 0.1.6 introduces delegated subagent execution with fanout orchestration, a runners dashboard/API surface, and stronger visibility controls across HTTP, Telegram, and TUI flows.
New Features
Delegated Subagent and Fanout Execution
New delegated-run tooling adds `subagent_spawn`, `subagent_fanout`, `subagent_status`, and `subagent_cancel` workflows for background and parallel subagent orchestration.
Delegated runs now track lineage and completion through a dedicated runner registry with return routing, cancellation flow, and persisted run metadata.
Multi-Provider Web Search and Better Telegram/HTTP UX
GoClaw now supports multiple `web_search` backends with retry/fallback controls, improves Telegram delivery for long and structured responses, and refines HTTP chat tool-panel behavior.
New Features
Multi-Provider `web_search`
`web_search` now supports `grok`, `brave`, `perplexity`, and `gemini` via a shared provider interface, including provider-aware result metadata.
Configuration and setup forms now expose provider selection, fallback ordering, retry tuning, and per-provider API key fields, while preserving legacy Brave key compatibility.
Telegram output handling was improved for long responses and structured tool activity display, with better delivery behavior and link handling for table-heavy content.
A regression affecting delivery of long Telegram messages was fixed after the initial rollout.
v0.1.4
HTTP Chat Rewrite, Binary Content Guard, and Safer Runtime Defaults
GoClaw 0.1.4 delivers a major HTTP chat/transcript UI rewrite, adds binary-content guardrails across tool outputs, and improves safety/runtime behavior for shutdown phrases, unknown model context windows, and headless environments.
New Features
Binary Content Guardrails for Tool Results
Tool outputs are now sanitized when binary or oversized content is detected, reducing context corruption risk from PDFs, archives, media bytes, or malformed stdout/stderr payloads.
Gateway settings now expose parallel execution controls with a safe allowlist model and max-concurrency tuning for lower-latency read/search style tool batches.
The HTTP channel UI was extensively rewritten for better chat usability, transcript handling, and frontend architecture, with updated docs that match the new workflow.
Sandbox Presets, Session Model, and Tool/API Renames
GoClaw docs now reflect the new setup wizard security presets, updated macOS sandbox behavior, and standardized tool/config naming across memory, transcript, and provider settings.
New Features
Wizard Security Presets
Setup now documents Assistant, Permissive, Hardened, and Custom security presets, including explicit acknowledgment requirements for preset-based paths.
Provider examples and field references now use `driver` instead of `type`.
Migration: Update `goclaw.json` under `llm.providers`: replace `type` with `driver` for each provider.
Transcript Tools Unified Under `transcript`
`transcript_search` and `transcript_stats` are replaced by action-based usage on the `transcript` tool.
Migration: Switch transcript tool calls to `transcript` and set `action` (for example `search` or `stats`).
Memory Graph Tool Names Prefixed
Memory graph tool names are now explicit (`memory_graph_recall`, `memory_graph_query`, `memory_graph_store`, `memory_graph_update`, `memory_graph_forget`).
Migration: Update role/tool allowlists and any tool-call references to the new `memory_graph_*` names.
Legacy Memory/Web Config Keys Replaced
Legacy `memorySearch` and `tools.web.useJina` config patterns are replaced by `memory.query` and browser fallback settings.
Migration: Move semantic search settings under `memory.query`; replace `tools.web.useJina` with `tools.web.useBrowser` and optional `profile`/`headless`.
v0.1.3
xAI Video Generation Tool Documentation
GoClaw now documents first-class `xai_video` support for text-to-video, image-to-video, and video editing workflows, including configuration, limits, and delivery behavior.
New Features
xAI Video Tool (`xai_video`)
A new tool page covers Grok video generation with prompt-only creation, still-image animation, and source-video editing, plus aspect ratio, resolution, timeout, and media-delivery options.
Installer Root Safety Guardrails and Sandbox Policy Clarity
The installer now adds explicit safeguards for root installs and improves PATH handling, while sandbox docs clarify expected hidden-path behavior in autodocs modes across Linux and macOS.
Improvements
Root Install Confirmation in Installer
Installing as root now requires explicit confirmation, with a clear warning about security risk and a `--allow-root` override for intentional non-interactive root installs.
Sandbox docs now explicitly state that autodocs hidden-path restrictions apply consistently to managed exec, managed browser launchers, and file tools.
Linux Hidden-Path Deny Behavior Explained
Docs now clarify that hidden-path access denial may surface as filesystem errors in exec while file tools return policy errors, and both outcomes are expected.
v0.1.3
Installer Dependency Controls and Better Linux Setup UX
The install script now supports explicit dependency handling modes and behaves better in non-interactive environments, making Linux installs more predictable in both manual and curl-pipe flows.
Improvements
New Dependency Mode Flags in Installer
The installer now supports `--deps auto|install|skip` plus `--yes` and `--no-deps` aliases so dependency installation behavior can be controlled explicitly.
Dependency confirmation now reads from `/dev/tty` so piped installs do not consume script stdin unexpectedly, and auto-skips cleanly when no interactive terminal is available.
Installer dependency installation now handles root/sudo detection more clearly and prints actionable manual commands when auto-install is not possible.
Sandbox docs now clarify that home-root listing can be used for discovery in macOS autodocs modes while hidden paths remain denied by policy.
v0.1.2
macOS Support & Built-in Embeddings
GoClaw now runs natively on macOS with full sandbox support. Semantic search works out of the box without requiring Ollama - a built-in embeddings provider handles it automatically.
New Features
macOS/Darwin Support
Full support for macOS including sandbox-exec (Seatbelt) for secure command execution. Build from source with Homebrew GNU Make (gmake).
Semantic search now works without external dependencies. The hugot-local provider automatically downloads and caches the all-MiniLM-L6-v2 model on first use.
Configure GoClaw through a browser-based interface in addition to the terminal wizard.
Improvements
Sandbox Autodocs Modes
New autodocs-read and autodocs-write modes expose common user directories (Documents, Desktop, Pictures) to sandboxed commands while keeping sensitive paths like ~/.ssh protected.