Eigent Release Notes v1.0.2: Connector Gateway, In-Session Preview & BYOK Controls

Eigent v1.0.2 is a release about keeping your work connected, visible, and moving. The new Connector Gateway brings hosted third-party app connections directly into Eigent through MCP, while the in-session Preview Panel gives browser pages, files, canvases, terminals, and reviews one place to live beside the conversation.

This release also gives BYOK users direct control over model parameters, makes sharing screenshots as easy as pasting into chat, clarifies what agents are doing in the live work log, and hardens the long-running workflows that make Eigent a true open source cowork platform. Let's dive in.

---

🔌 Connector Gateway: Your Apps, Available to Your Agents

The biggest new capability in v1.0.2 is a direct path from Eigent to the tools where your work already lives.

Huge thanks to @4pmtong for building the first Eigent-side Connector Gateway integration in PR #1776.

The Connector Gateway adds a provider-agnostic interface for hosted connector runtimes. Instead of wiring every service directly into the desktop app, Eigent connects through the hosted eigent_server, receives the appropriate MCP configuration, and makes those tools available when a new task starts.

What's new:

• Connector discovery — browse available providers with search and pagination from the Connectors area
• Connector detail drawer — inspect provider metadata, connection state, available actions, and action previews before using a connector
• Capability-aware UI — the Connector Gateway only appears when the connected Eigent server advertises support
• Automatic MCP mounting — connected tools are added to new tasks through the Gateway's MCP configuration
• Connector-aware agent prompts — agents now prefer the relevant MCP tools when you ask them to query or operate a connected third-party app
• Provider-agnostic architecture — the integration is named and structured so Eigent is not permanently coupled to one connector provider

This is an important step toward a cowork environment where agents can move from planning to action without forcing you to copy information between apps by hand.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1776

---

🖥️ One In-Session Preview Panel for Every Kind of Work

Your work can now stay beside the conversation instead of opening in scattered windows.

Shoutout to @Douglasymlai for introducing the unified Preview Panel in PR #1750.

The old single-purpose file preview has become a tabbed workspace for browser, file, canvas, terminal, and review content. It sits in a resizable split next to chat, opens and closes with a smooth transition, and remembers its state separately for each project.

What's new:

• Five preview types — open browser, file, canvas, terminal, and review tabs in one consistent panel
• Embedded desktop browser — navigate with back, forward, refresh, an address bar, and an option to open the page in your system browser
• Links stay in context — clicking an HTTPS link in a desktop chat opens it inside the session preview instead of pulling you away from Eigent
• Resizable session layout — adjust the balance between conversation and preview as the task changes
• Per-project preview state — tabs remain scoped to the right project when you switch between sessions
• Safer webviews — unsafe guest preferences and non-HTTP(S) sources are blocked, while new-window links remain inside the controlled preview
• Resource-aware behavior — idle browser guests are evicted after ten minutes to reclaim renderer resources

Whether an agent has generated a file, opened a research source, or produced something that needs review, the result now has a natural home inside the session.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1750

---

🎛️ BYOK Model Parameters and Session-Level Model Choice

Bring your own key now also means bring your own inference settings.

Thanks to @fengju0213 for adding BYOK model parameters in PR #1765, and to @iyernaveenr for making model selection persist per chat session in PR #1736.

Previously, BYOK users could select a provider and model but had no supported interface for settings such as temperature, top_p, or max_tokens. Eigent v1.0.2 adds an optional JSON editor to every BYOK provider card, validates the object inline, saves it with the provider, and forwards it correctly to both the default task model and provider-backed worker overrides.

What's new:

• Model Parameters editor — configure supported inference parameters as a JSON object from the BYOK provider card
• Immediate validation — malformed JSON and non-object values are caught before they reach the model
• Settings that round-trip — parameters are encrypted, persisted, and restored when you reopen the provider
• Worker-level propagation — custom and local workers receive the selected provider configuration without storing API keys or provider secrets in worker state
• Safer logs — per-worker model credentials are redacted from new-agent debug output
• A model for each session — existing conversations remember their own selected model across follow-ups and app restarts, without changing the global default

The result is more precise control for advanced users and more predictable conversations when several projects use different models at the same time.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1765

🔗 PR: https://github.com/eigent-ai/eigent/pull/1736

---

📎 Paste Images into Chat and Trust the Files You Get Back

Sharing visual context with an agent should be as quick as sharing it with a teammate.

Thanks to @iyernaveenr for clipboard file paste support in PR #1740, and to @4pmtong for the output-file and Agent Folder fixes in PRs #1767, #1773, and #1798.

You can now copy a screenshot and paste it directly into the chat box. On desktop, Eigent securely saves the clipboard bytes to a temporary file and passes the result into the existing attachment flow. On the web, pasted files use the same upload path as dropped files. Plain text still pastes normally, even when the clipboard contains both text and an image.

What's improved:

• Clipboard image paste — use Cmd+V or Ctrl+V to attach screenshots without opening a file picker
• Clipboard file support — file items copied from a file manager can enter the same attachment pipeline
• Consistent safeguards — paste follows the same availability rules, deduplication, and success or error feedback as drag-and-drop
• Correct sandbox paths — sandbox:/... output paths no longer get misread as Windows drive paths
• Safer file detection — URLs and unknown schemes are not presented as local output files
• User-scoped output resolution — files are resolved from the current user ID path, with legacy email-based storage retained as a fallback
• Cleaner Agent Folder — internal task-root directories are hidden and directory entries are no longer shown as clickable files

These changes close both sides of the loop: it is easier to give an agent visual input, and more reliable to open the files the agent creates.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1740

🔗 PR: https://github.com/eigent-ai/eigent/pull/1767

🔗 PR: https://github.com/eigent-ai/eigent/pull/1773

🔗 PR: https://github.com/eigent-ai/eigent/pull/1798

---

🔎 Clearer Work Logs, Backed by Typed Workflow Events

When agents are working, Eigent should make their progress easy to understand and dependable to process.

Thanks to @Pakchoioioi for the work-log improvements in PRs #1696 and #1697, @Douglasymlai for the typed app event bus in PR #1771, and @fengju0213 for request-level token usage reporting in PR #1362.

Tool activity now separates the request sent to a tool from the response that came back. In Single Agent mode, the work-log header also replaces the generic “CAMEL Agent” label with the live active_form for the step currently in progress, updating with a subtle transition as the agent moves through its plan.

What's changed:

• Request and Response cards — tool inputs and outputs are labeled separately instead of being merged into one block
• Live step names — Single Agent work logs show the current action and update as work advances
• Long labels wrap cleanly — active step text remains readable instead of being truncated
• Typed app event bus — workflow paths emit structured app events through a lightweight, testable foundation
• Early-event buffering — startup events wait safely until the consumer is ready
• Request-level usage events — non-streaming, multi-request steps can report token usage as each model request completes instead of waiting for the entire step
• Double-count protection — request-level and step-level usage reporting are coordinated so tokens are not counted twice

These are user-facing clarity improvements with a stronger event foundation underneath—useful today and ready for richer workflow diagnostics tomorrow.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1696

🔗 PR: https://github.com/eigent-ai/eigent/pull/1697

🔗 PR: https://github.com/eigent-ai/eigent/pull/1771

🔗 PR: https://github.com/eigent-ai/eigent/pull/1362

---

✨ A More Thoughtful Chat and Desktop Experience

The small interactions around a task now feel calmer, clearer, and more deliberate.

Thanks to @Douglasymlai for redesigning the chat composer, tooltips, update experience, support-log flow, and motion system in PRs #1749, #1754, #1751, #1802, and #1805. Thanks as well to @iyernaveenr for restoring the last visited session at launch and opening switched chats at their latest message in PRs #1745 and #1733.

What's improved:

• Two-tier chat composer — message actions sit above a dedicated project-setup row for session mode, model, and thinking effort
• Direct connector and skill pickers — add @connector and #skill tokens without routing through a generic plus menu
• Intent-aware tooltips — navigation labels appear instantly, while richer informational hints keep a short, calmer delay
• Inline update progress — desktop updates download in the background and show progress in the top bar, followed by a clear “Launch new version” action
• Downloadable diagnostic logs — the Support dialog can export both Eigent desktop logs and CAMEL task logs
• Resume where you left off — reopening the app restores the last visited conversation when appropriate
• Latest message on project switch — changing chats takes you straight to the current end of the conversation
• More consistent motion — targeted transitions replace broad transition-all behavior across onboarding, tasks, workflows, skills, connectors, sidebars, and dialogs, with reduced-motion support where appropriate

Individually, these are compact changes. Together, they remove friction from the parts of Eigent you touch all day.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1749

🔗 PR: https://github.com/eigent-ai/eigent/pull/1754

🔗 PR: https://github.com/eigent-ai/eigent/pull/1751

🔗 PR: https://github.com/eigent-ai/eigent/pull/1745

🔗 PR: https://github.com/eigent-ai/eigent/pull/1733

🔗 PR: https://github.com/eigent-ai/eigent/pull/1802

🔗 PR: https://github.com/eigent-ai/eigent/pull/1805

---

🛡️ Long-Running Workflows That Stay Running

Reliability matters most when a task has already been working for a long time.

Major thanks to @iyernaveenr for protecting live runs across project switches and window closes, making step timeouts configurable, recording the spend of failed runs, and fixing completion state in PRs #1748, #1747, #1746, and #1737. Thanks to @4pmtong for strengthening background remote control, backend startup, Space hydration, and stale reply handling in PRs #1774, #1775, #1797, and #1799.

What's more reliable:

• Project switches preserve live runs — moving between projects no longer evicts or replays state in a way that can abort an active stream
• Close warnings cover every project — Eigent warns before the window closes while any run is active, not only the task currently on screen
• Configurable step timeout — deployments can change AGENT_STEP_TIMEOUT_SECONDS, retain the 30-minute default, or disable the per-step ceiling
• Failed-run spend is recorded — token usage accumulated before a timeout or error is no longer left at zero
• Long answers still complete — history summaries are clamped to the database limit without preventing the task status from becoming done
• Background Remote Control works — remote messages can target a background project without forcing the desktop UI to switch projects
• Safer local backend startup — duplicate launches, restart races, and broken console pipes are handled more defensively
• Active Space survives hydration — server synchronization preserves the user's selected Space and only creates a blank one for genuine setup or migration cases
• Late human replies fail gracefully — stale replies after task-lock cleanup clear outdated UI state instead of surfacing as backend program errors

For anyone using Eigent on long research, coding, or document workflows, these fixes make the desktop app a much safer place to leave serious work running.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1748

🔗 PR: https://github.com/eigent-ai/eigent/pull/1747

🔗 PR: https://github.com/eigent-ai/eigent/pull/1746

🔗 PR: https://github.com/eigent-ai/eigent/pull/1737

🔗 PR: https://github.com/eigent-ai/eigent/pull/1774

🔗 PR: https://github.com/eigent-ai/eigent/pull/1775

🔗 PR: https://github.com/eigent-ai/eigent/pull/1797

🔗 PR: https://github.com/eigent-ai/eigent/pull/1799

---

🐛 More Fixes Across Sign-In, Chat, OAuth, Windows, and Docker

This release also includes targeted fixes across the stack:

• Fixed local sign-in losing the user account state — local sessions recover the user ID from the login response or access token, preventing an empty read-only Legacy Space from replacing the user's real Spaces and Projects. (#1739 — @iyernaveenr)

• Fixed invisible fenced code blocks when OS and app themes disagree — code text and background colors now follow the same app theme, keeping all four light/dark combinations readable. (#1735 — @iyernaveenr)

• Fixed the chat caret jumping during fast typing — caret restoration now happens synchronously after rich-text rendering, while only the layout-dependent scroll waits for the next frame. (#1732 — @iyernaveenr)

• Fixed Single Agent skill scoping — Single Agent is always visible first in the skill access picker, and its aliases are normalized consistently across frontend and backend matching. (#1759 — @suger-m)

• Fixed Windows terminal capability detection — native PowerShell and Command Prompt environments can register terminal tools even when Bash is not installed. (#1716 — @fengju0213)

• Fixed backend trust-store handling — bundled desktop backends can use operating-system-trusted certificates, including private CAs used by remote HTTPS model endpoints. (#1700 — @fengju0213)

• Fixed OAuth authorize URL encoding — Slack, Notion, X, and Google adapters now encode redirect URLs, scopes, state, Unicode, and other special characters correctly. (#1636 — @hobostay)

• Fixed fresh Docker server builds — .git remains available to the BuildKit bind mount used to stamp the server commit, without being copied into the image layer. (#1703 — @nilbot)

---

📚 Dependencies and Documentation

Thanks to @fengju0213 for upgrading camel-ai to 0.2.91a5 in PR #1760, @carlosjarenom for correcting the server README's Python requirement from 3.10 to 3.12 in PR #1672, @Douglasymlai for refreshing the product highlights and use cases across the localized READMEs in PR #1802, @4pmtong for separating open source build uploads into their own S3 prefix in PR #1801, and @Wendong-Fan for updating the WeChat QR code.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1760

🔗 PR: https://github.com/eigent-ai/eigent/pull/1672

🔗 PR: https://github.com/eigent-ai/eigent/pull/1801

---

❤️ Community Keeps Raising the Bar

This release delivered:

• A Connector Gateway that makes hosted third-party tools available to agents through MCP
• A unified in-session preview for browser, file, canvas, terminal, and review work
• BYOK model parameters and model choice that persists per conversation
• Clipboard image paste and more dependable generated-file handling
• Clearer live work logs, typed workflow events, and request-level token reporting
• A redesigned composer, smarter tooltips, visible update progress, and downloadable support logs
• Stronger protection for long-running tasks across project switches, window closes, timeouts, and background control
• Focused fixes across sign-in, chat rendering, skills, Windows, certificates, OAuth, and Docker

Eigent v1.0.2 is the product getting better at the work between the big moments: connecting the right tool, keeping context visible, explaining what an agent is doing, and protecting a task while it runs.

Every feature and fix came from contributors turning real workflow friction into a better open source cowork experience. If you want to help shape what comes next, explore the code, open an issue, or send a PR.

🔗 Release: https://github.com/eigent-ai/eigent/releases/tag/v1.0.2

Let's keep building.