Eigent Release Notes v0.0.90: Prompt Caching, Browser File Upload & Faster Updates

Eigent v0.0.90 is a release about doing more with less — and doing it faster. The headline feature is platform-wide prompt caching: instead of re-processing the same context on every agent turn, Eigent now caches it at the model level, slashing latency and cutting costs significantly for long-running workflows. Alongside that, the browser agent gains file upload capability, and the auto-update pipeline moves from GitHub Releases to a CDN for dramatically faster downloads.

For anyone running Eigent as their open source cowork platform, v0.0.90 is a release that quietly makes everything feel snappier and more capable. Let's dive in.

---

⚡ Prompt Caching for All Models

Performance and cost just took a major step forward.

Huge thanks to @Zephyroam for the foundational implementation in PR #1482, and to @fengju0213 for the targeted follow-up in PR #1552 that extends caching to Anthropic BYOK model initialization.

Prompt caching allows the model to store and reuse the processed representation of long, stable context — system prompts, skill instructions, document context — rather than re-tokenizing and re-processing it on every agent turn. For the kind of multi-step, long-horizon tasks that Eigent cowork agents excel at, this can mean dramatic reductions in both latency and token costs.

PR #1482 is a broad implementation that extends prompt caching support across multiple model platforms, with particular depth on AWS Bedrock Converse:

What's new:

• AWS Bedrock Converse integration — dedicated prompt caching support for Bedrock Converse with configurable region handling and support for both API key and access key/secret key authentication
• AWS Bedrock vs. Bedrock Converse separation — the two are now treated as distinct platform types, since Bedrock Converse supports prompt caching while standard Bedrock uses OpenAI-compatible APIs that do not
• Cloud-specific configuration — Bedrock Converse models get proper URL and region-aware configuration at initialization
• Improved validation error messages — cleaner, more descriptive errors across different platform types to help users debug BYOK configuration issues
• Agent worker stream callbacks — aligned with upstream CAMEL framework changes for proper streaming behavior

PR #1552 (by @fengju0213) closes the loop by adding cache configuration at the Anthropic model initialization stage, ensuring that Anthropic BYOK users also benefit from prompt caching without any manual configuration.

In a multi-agent cowork workflow where agents share a long system prompt or skill document, caching that context once instead of processing it on every call is the kind of optimization that compounds across every task.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1482

🔗 PR: https://github.com/eigent-ai/eigent/pull/1552

---

🌐 File Upload Tool for the Browser Agent

The browser agent just got a powerful new capability.

Shoutout to @Douglasymlai, reviewed by @nitpicker55555, for adding file upload as a first-class tool in the browser agent. Previously, the browser agent could navigate, read, and interact with pages — but couldn't handle file upload inputs. That gap is now closed.

Web automation frequently encounters file upload scenarios: attaching a document to a form, submitting a spreadsheet to a data portal, uploading an image to a CMS. Without native upload support, these workflows required awkward workarounds. Now the browser agent can handle them cleanly.

What's new:

• Upload File tool — a dedicated tool in the browser agent toolset for handling file input elements on web pages
• End-to-end upload support — the agent can select a file and trigger upload interactions as part of a larger browser automation workflow
• Integrated with existing browser toolchain — works seamlessly alongside existing navigation, form input, and reading tools

This makes Eigent's browser agent significantly more capable for real-world open source cowork automation scenarios that involve file submission, document intake, or media upload.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1528

---

📊 Auto-Scroll Worker Log

Small UX change. Big difference when you're watching a task run.

Thanks to @Douglasymlai, reviewed and approved by @4pmtong, for implementing automatic scrolling in the workflow worker log panel.

Previously, when an agent was executing a long task with many tool-use entries, the log would fill up and new entries would appear below the visible area. Users had to manually scroll down to follow along — disruptive when you're actively monitoring a cowork agent session.

What's new:

• Auto-scroll on new entries — the log view stays pinned to the latest tool-use message as new content arrives during task execution
• Scroll state tracking — a wasAtBottomRef tracks whether the user is viewing the bottom of the log, so auto-scroll doesn't interrupt if they've intentionally scrolled up to review earlier steps
• Reset on task switch — the at-bottom state resets when switching tasks or opening the log panel, ensuring a clean start for each new agent run
• Layout refinements — updated Tailwind class strings in the workflow node component for improved panel layout

When you're running multi-step eigent cowork agents and want to see what's happening in real time, this change makes the log panel feel like a proper live feed.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1513

---

⚙️ Stale Server Detection for Local Development

Never wonder "am I running old code?" again.

Thanks to @4pmtong for building and shipping this developer experience improvement.

Anyone who runs Eigent with VITE_USE_LOCAL_PROXY=true has hit this: you pull new changes, forget to restart the backend, and spend time debugging behavior that was already fixed — because the server is running stale code. PR #1517 solves this with automatic detection and a persistent notification.

What's new:

• Git hash at startup — the backend now captures SERVER_CODE_HASH from the current git commit at server start
• Extended /health endpoint — now returns both version and server hash, making it possible for the frontend to compare expected vs. actual server state
• checkLocalServerStale() function — the frontend polls the health endpoint, compares hashes, and fires a persistent toast notification when a mismatch is detected
• Git hash vs. version number — intentionally uses git hash rather than version string, to avoid false positives on frontend-only releases that don't change the backend
• Docker support — the Dockerfile is updated so the hash is available inside containerized dev environments too

If you contribute to eigent or run it locally for your open source cowork workflows, this is a small but genuinely welcome quality-of-life improvement that will save debugging time.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1517

---

📦 CDN-Powered Auto-Updates

Faster upgrades. Less waiting.

Thanks to @4pmtong, reviewed by @Pakchoioioi, for switching the auto-update feed from GitHub Releases to a CDN.

Previously, Eigent's Electron auto-updater pulled installation packages directly from GitHub Releases. This worked, but GitHub's release infrastructure isn't optimized for binary distribution speed at scale. The switch to a CDN feed means update packages are now served from edge nodes closer to users.

What's changed:

• CDN as the update source — the Electron updater now points to a CDN-hosted feed instead of GitHub Releases for pulling packages
• Faster download speeds — especially noticeable for users outside North America where GitHub CDN edge coverage is thinner
• Same release cadence — nothing changes about when releases ship; only where the packages are downloaded from

If you've noticed slow update downloads when Eigent prompts you to upgrade, this change should make that experience significantly better.

🔗 PR: https://github.com/eigent-ai/eigent/pull/1543

---

🐛 Bug Fixes

This release also includes three targeted fixes:

• Fixed NameError on backend startup with public directory warnings — server/main.py was referencing an undefined logger variable when logging public directory creation or mount warnings, causing a NameError during initialization. Fixed by replacing it with the correctly imported loguru_logger. (#1520 — @mxl)

• Fixed browser port error reopen issue — resolved a bug causing the browser agent's port to fail on reopen, which could leave the browser tool in an unrecoverable state mid-session. (#1545 — @Douglasymlai)

• Fixed Anthropic model validation rejecting missing max_tokens — the Anthropic Messages API requires a positive integer max_tokens value, but the BYOK form doesn't expose this field. When validation ran without it, Anthropic returned a 400 error. The fix ensures validate_model_with_details and create_agent both merge a default of 4096 for the max_tokens field when missing on Anthropic platform. (#1549 — @emag165, refined by @fengju0213)

---

❤️ Community Keeps Raising the Bar

This sprint delivered:

• Platform-wide prompt caching for all model providers, with depth on AWS Bedrock Converse and Anthropic BYOK
• File upload capability added to the browser agent
• Auto-scroll in the worker log for live task monitoring
• Stale local server detection for developers running with a local proxy
• CDN-backed auto-update downloads for faster upgrades
• Three targeted bug fixes across backend startup, browser port handling, and Anthropic model validation

Every one of these changes came from contributors who use Eigent, hit a real problem, and pushed a fix or feature. That's the rhythm of open source cowork at its best.

If you've been thinking about contributing — whether that's models, UX, agent tooling, skills, or infrastructure — v0.0.90 is a great snapshot of what's possible. Jump in.

Let's keep building.