Inspect, organize, and safely maintain coding-agent context, Skills, sessions, and local toolchains.
Coding agents build up useful local state across project instructions, Skills, MCP connections, native configuration, and conversation history. That state is fragmented between different tools, making it hard to know what an agent can actually see, reuse trusted assets, or continue work safely elsewhere.
AgentKib brings that state into one local, inspectable desktop surface. Core discovery, diagnostics, Skill management, and session handoff do not require an AgentKib account, cloud database, or model API. KIB stands for Knowledge & Instruction Base.
See the Instructions, Skills, MCP connections, memory, and native configuration available to each agent in a workspace. Diagnose missing, drifted, invalid, or duplicated assets, then review a complete ChangeSet and diff before anything is written.
Browse reviewed OpenAI Skills or inspect a public GitHub repository before adding a package to the local library. AgentKib pins managed packages to immutable commits, previews files and executable resources, detects updates and local drift, and supports rollback and recoverable removal. Library Skills are not enabled for an agent automatically.
Inspect and copy existing installations into the library, then review deployments to personal or project locations. Each deployment can be updated, rolled back or withdrawn independently without taking over the original installation. Shared visibility and native loading restrictions are reported separately. See Skill management.
Desktop and Web also offer opt-in conversation content search on the current host. Find message or readable tool text, inspect a passage, and quote it into the current draft without switching the target conversation. Only an explicit send includes the quote. Other paired hosts remain title-search only. See content search and source coverage.
Browse Codex, Claude Code, Antigravity ACP, OpenCode, supported Cursor CLI stores and explicitly connected Cursor IDE 3.22.12/3.23.12 profiles, OpenClaw schema-23 SQLite and legacy JSONL, Hermes, and Grok Build session history. Reviewed handoff availability is checked against the selected native history, including format and completeness checks. OpenCode 1.18.32, Hermes 0.21.5, and OpenClaw 2026.9.6 have experimental native-import adapters with durable recovery records; offline import/readback is verified, with OpenCode and Hermes also checked in their native terminal UIs after restart. September 30 Claude-to-OpenCode and Codex-to-OpenCode text cases each completed one real reply through the user-selected CPA/Opus provider, correctly recalling imported context and preserving the same session after Runtime restart. This does not validate every source, tool record, attachment, or other target. One OpenClaw-to-Claude synthetic history has produced a correct real reply; full per-direction acceptance is still pending. See the direction and version matrix. AgentKib preserves useful timeline context, redacts common sensitive values, and requires confirmation before writing or importing handoff artifacts. For a compacted Claude Code session, the default handoff carries the compaction summary plus the messages after it and declares the earlier history it leaves out; an advanced option adds that earlier history. Antigravity's current native-import and Desktop/CLI interoperability limits are documented in Antigravity integration.
The October 1 offline acceptance closeout records all 35 representative text directions across the five existing targets. Claude native TUI recovery and Codex official-interface recovery are verified for seven available sources. The subsequent provider retest verifies two independent offline OpenClaw TUI restores for each of its six remaining sources and a real Codex-to-Claude reply using the newly selected DeepSeek provider. Exact source, target identity and full-history checks remain separate from model-answer acceptance; unsuccessful replies and unverified release gates remain open.
The October 1 closeout adds an exact Codex 0.155.1 compatibility gate while preserving the 0.146 contract. New Grok→Codex 0.146.1, Claude→Codex 0.155.1, and Grok→OpenClaw 2026.9.6 cases each made one DeepSeek request and recovered the full native history twice offline. Original failed results remain intact; the OpenClaw punctuation difference has a separate exact full-sentence audit. Cursor IDE 3.22.12 now has an experimental local first-party VSIX bridge, explicit profile/window pairing, reviewed native import, full readback, durable recovery, and a separate source namespace. Two isolated Claude→Cursor IDE text cases pass native storage import, full content checks, same-ID recovery, and visible history after official login. The first real reply failed context inheritance: the converter populated UI history but omitted separate model prompt history. That defect is fixed with strict readback regressions. Two corrected cases independently verify all five model/UI messages and thirteen blobs, with complete visible history and unique native identities. One case recovers its original operation after vendor protobuf field grouping and reopens the same ID after two Runtime restarts; the other imports the canonical format directly. A new B case made one GUI submission and correctly recalled the complete historical marker and storage decision. Both histories remained visible after a second normal client restart. Post-reply recovery exposed a fixed-version assistant JSON key-order change; the narrow compatibility fix passes tests and actual same-operation reopening after a Runtime restart. October 2 checks also preserve all seven messages when this native history becomes a source again for five target previews; these previews do not represent five native imports. The old failures remain intact. Cursor IDE→Claude also passes one native file import and two offline native TUI restores, without a model call. CLI and Agents Window capabilities remain separate. Actual Cursor results are recorded independently from fixtures in the native acceptance QA and closeout QA.
These historical native results apply to the source revisions, Agent versions, and directions recorded in each QA report. The current TypeScript backend has separate stability and packaging acceptance; historical Rust Runtime or model-reply evidence does not certify the migrated backend.
Claude Code also has an opt-in native-session bridge for the fixed macOS CLI 2.1.301. The first-party Mods plugin connects the existing executor to the desktop and Web conversation UI; it does not silently start a second resume process. Observation and current-turn cancellation are enabled. Text sending stays disabled because atomic no-queue admission is unverified; approvals, questions and attachments remain separate from managed execution. See native Claude connection and its current QA boundaries.
Collect native or pasted MCP configurations without starting services, explicitly enable and probe them, and review one batch to connect selected installed Agents to the Hub. Assign individual tools per Agent, with workspace overrides and checked configuration writes. See MCP management.
Inspect Codex, Claude Code, Antigravity, Cursor, OpenCode, OpenClaw, Hermes, and Grok Build installations. AgentKib reports each installation's version, source, executable path, PATH default, and conflicts. Verified package-manager actions can run with a pinned target; channels that cannot pin an exact version, along with ambiguous, privileged, or remote-script flows, fall back to a command or official documentation.
This monorepo includes a separately built Web client bundled with Electron. In Settings → Remote connections, use the phone-access setup action and sign in to your account and complete device registration and pairing. The default registration API is https://api.agentkib.com; https://remote.agentkib.com remains the hosted Web entry and trusted LAN origin. Local/LAN Web access is a separate, optional route that does not require the hosted service. Native connections to another desktop have their own pairing flow; see native LAN connections. The desktop application must remain running. Read access is separate from control permissions. Claude Code uses the same managed-session backend for desktop and Web: new sessions, confirmed same-UUID continuation, images/files, approvals, and durable receipt recovery. Basic control retains its macOS compatibility gate; advanced control requires Claude Code 2.1.286 or later, host enablement, and the relevant device permissions. Native and device acceptance is tracked separately in the historical Claude QA and current parity QA; Codex owner control remains acceptance-gated. See the local Web setup guide (Chinese), Claude control boundaries, and QA status. This preview is not a claim that the current published release includes Web access.
The Claude Remote implementation now includes discovered model/effort settings, session-scoped tool permissions, native next steering, verified native-history forks and native titles, plus AgentKib-managed queues, archives and goals. Model-only changes preserve CLI-inherited permissions. AgentKib still passes no --permission-mode unless you pick one, and shows the mode the CLI actually reports: auto, dontAsk and bypassPermissions get a visible notice in the desktop and Web session views and are never offered as choices. A mode the CLI never reports stays unknown, with a "may be auto" hint on Claude Code 2.1.285 or later, which can start in auto when it cannot fetch feature flags (for example with a third-party provider or telemetry off); this is verified against synthetic frames only, not a real CLI. Effort changes recognize native model aliases, and resetting effort alone restores CLI configuration while retaining the model and permission selections. Unstarted forks retain their source and cutoff across release and confirmed handoff. Queues and goals require explicit resumption after a host/backend restart; goals use native turn usage for budgets and require both a structured MCP report and the matching successful native result before completion. Missing usage remains recorded across updates and restarts: bounded goals cannot resume until actual usage is recovered or the user explicitly removes the budget. Queue editing is limited to pending text entries and preserves the original device; dispatched entries and attachment references stay fixed. The browser keeps HTTP commands and SSE updates; an optional /api/web/v1/socket WebSocket adapter uses the same authorization and receipts for bounded requests. Advanced controls and this socket are not enabled for hosted-Web-to-LAN HTTP connections. Automated results do not establish real-model, physical-phone or production-relay acceptance.
The remote implementation adds Codex app-server task creation and explicit same-ID handoff, separately granted project files/artifacts and read-only Git diffs, isolated interactive HTML bundles, and streamed media previews. AgentKib Remote account registration is open at account.agentkib.com: a username and password are sufficient, with no registration invitation code and a default limit of three devices for new accounts. Existing account limits, device authorization, pairing, and host/browser control permissions remain separate. The server implementation and deployment configuration are not included in this repository; pricing is not announced here. Previously configured https://remote.agentkib.com broker identities require explicit migration and new registration; credentials are not copied to the new API origin. The desktop connector keeps browser business TLS termination on the desktop. This repository contains the desktop connector, Web client and local agent/artifact services. See the Codex and artifact setup and hosted remote service guide. The deployed registration page and API health have been checked, but these checks do not establish full production account-flow or physical iOS/Android acceptance. See the current QA boundaries.
Download the current stable package from the Latest Release:
- macOS 13.3+:
.dmgfor Apple Silicon or Intel - Windows 11: x64
.exe; ARM64 remains Preview - Linux: Ubuntu
.debor AppImage, or Fedora.rpm; ARM64 remains Preview
Only use files from the official release and verify the matching .sha256 checksum. Windows installers are not yet Authenticode-signed and may trigger SmartScreen. See Upgrading for in-app updates, package-specific upgrade paths, and historical macOS packages.
- Ordinary discovery and diagnostics do not invoke a model or upload project assets.
- Credentials, cookies, tokens, private keys, environment files, and message databases are not cataloged as assets or written to logs.
- Claude Code managed execution starts your own
claudeCLI, which keeps using its own sign-in. AgentKib does not read, copy, or issue Claude login credentials for it. - Quota collection is a separate path run by a bundled third-party sidecar (CodexBar), and only on a manual refresh or after you turn on "Automatically get quota" (off by default). AgentKib itself does not read provider credentials, but the sidecar may: for Claude it can read claude.ai browser cookies, run the
claudeCLI, use Claude Code's OAuth credentials if that source is selected in the CodexBar configuration, and scan local Claude Code logs. See quota collection and credentials. - Generated writes validate path boundaries and original hashes, create backups, and require review before applying.
- Agent Home changes use a separate high-risk confirmation flow; remote integrations retain their own permission boundaries.
- Discovery and context: Codex, Claude Code, Antigravity, Cursor, OpenCode, OpenClaw, Hermes, Grok Build, and read-only DeepSeek Harness diagnostics.
- Session browsing: Codex, Claude Code, Antigravity ACP, OpenCode, supported Cursor CLI stores and explicitly connected Cursor IDE 3.22.12/3.23.12 profiles, OpenClaw schema-23 SQLite and legacy JSONL, Hermes, and Grok Build. Reading never modifies the original history.
- Reviewed handoff: all eight sources above, subject to per-session parsing. Native import, context-file handoff, and Web control are separate capabilities; see the compatibility matrix for restrictions and acceptance status.
- Tool management: Codex, Claude Code, Cursor, OpenCode, OpenClaw, Hermes, and Grok Build. Antigravity is detected and linked to official update guidance, but has no automatic package action. DeepSeek Harness is intentionally excluded.
- Interface: English, Simplified Chinese, Traditional Chinese, and Japanese; light, dark, and system themes.
- Platforms: macOS 13.3+, Windows 11, Ubuntu 22.04, and Fedora. ARM64 Windows and Linux packages remain Preview.
See the complete feature and compatibility matrix for global views, workspace capabilities, agent support, and platform status.
For development-only read-only sources and missing-workspace troubleshooting, see the discovery and history guide (Chinese) and layered QA record. These additions are not a claim about released packages.
- Feature matrix
- Development guide
- Upgrade guide
- Release process
- Contributing
- Security policy
- Third-party notices
- GitHub Issues
Report security vulnerabilities privately according to the security policy, not through a public issue. AgentKib is released under the MIT License.