LiveContext Files

What is new

The release that makes a run readable: the inspector reports what each node actually ran with and which branch it took, run logs move into the side panel, and a parked run says what it is waiting for. It also brings a second engine for the Classify node, a mailbox tool for agents, password reset by e-mail for self-hosted installs, and a license change (see Self-hosting & install).

Workflows

• The run inspector reports what actually happened. Decision, switch, option and loop nodes now say which branch was taken, what the condition was and what it resolved to, instead of showing the same result on the branch that ran and the one that was skipped. Parameters come from the evaluation that decided, and a reference that points at nothing is shown as unresolved, with the reason.
• Every node shows what it ran with, without leaking secrets. Interface, split, find and paused nodes (waiting on a continue, an approval or a long wait) now report the values they resolved, and API keys no longer reach the Params column.
• Run logs live in the side panel. Log modals are replaced by a run view in the side panel with Run and Logs navigation, node-scoped logs, a workflow-then-run picker, and separate state for each panel so two open workflows stay isolated.
• A parked run says what it is waiting for. The run indicator is blue with a moving bar while the engine works, and amber and still while it waits on a person. A parked interface node offers Continue directly on the canvas, and an application shows an amber bar with the action instead of hiding it in the toolbar.
• The canvas camera can follow the running step. A new toggle beside Focus, in run mode, keeps the canvas framed on whatever node is currently running.
• Per-node control of provider retries. A node can cap how long the platform waits and retries when a provider answers "too many requests", so a workflow that paces itself no longer multiplies retries. Agents can set it when building workflows.
• Filter workflows and applications by node type, from the lists and from the agent list actions, so "which of my workflows use Gmail?" has an answer.
• Application trigger shortcuts. Application triggers can be added as shortcuts, keeping their node type and resolving their tools from the catalog.
• Run data is easier to read. Long strings collapse like any other row, the copy and expand controls no longer eat the value's width, the table view works for payloads with several fields, and embedded JSON renders like a real object.
• Download File follows redirects safely. Links such as Google Drive shares work again: each hop is validated before it is requested, with a limit of 5 hops, loop detection, and a refusal to downgrade from https to http.

Agenda

• Schedule an agent from an empty slot, not only a workflow. The calendar fills in the time and timezone, and the agent is created only when you save its form.
• Agent runs appear on the calendar, labelled with how they were launched.
• Trigger search, resource controls and clearer pause scopes, with compact dates and the calendar opening in the side panel. On a phone, the month view no longer shows a grid of clipped times.

Agents & chat

• A mailbox tool for agents. Agents can read, search and send e-mail through your configured mail accounts. It is off by default and bounded by four gates: the tool grant, a read or read-write access mode, the workspace role, and a confirmation card before sending or deleting. The general chat has its own mailbox switch, which starts in read-only mode.
• Confirmation before an agent arms something that outlives the chat. Pinning or unpinning a workflow version, and creating or updating an agent with a schedule, now show an authorization card naming the workflow and version, or the cron and timezone.
• Filter the model menu by tier and provider. The composer's model menu gains a footer that narrows the list by price tier and provider, and the reasoning-effort control moves there too.
• Configurable conversation compaction. Long conversations can be summarized by number of turns (the default, unchanged), by size, or by whichever comes first, through CONVERSATION_COMPACTION_TRIGGER.
• Attached files can be used in tool calls. An agent can pass a file attached in chat to a tool, for example as the source image of an image edit, instead of only seeing it.
• Smoother chat streaming. The end-of-stream refresh is invisible, a conversation no longer loses its live connection silently, and the collapsible reasoning section is back.
• Agent completion, memory and reasoning history improved, and the "go to agent" link opens the right page.

Models

• A second engine for the Classify node. Classify can now run on a decision model (TypeSafe Jev) that returns a typed choice among your categories with a probability for each. Same node, same categories, same ports: switching is only a provider and model change. Its API key is set on the AI Providers page like any other provider.
• Switch a whole AI provider off in one move. Disabling a provider hides it and all its models from every picker while keeping each model's own setting, so turning it back on restores your selection.
• Bulk edits and filters in the Models panel. Tick several models and change them at once, filter by name, tier and state, and see a renamed model's real id under its display name. Renames now stick after a catalog refresh.
• The bundled model catalog is refreshed, and a seed versioning issue that could stop it from applying is fixed.

Studio & generation

• First frame, last frame and references at once. Generation models that accept several images for different roles can now receive all of them, and rules such as "a last frame needs a first frame" are published with each model so the interface prevents the mistake instead of reporting it. Existing workflows keep working unchanged.
• Darkroom look for Studio, a second light or dark treatment picked from the composer and remembered per user.
• Add your own provider key from Studio for every video provider in the catalog.
• Generation cost is shown, and a generation is refused before it is charged when storage is full.

Tables & files

• Every media kind previews in an attachment cell, not just images.
• The column order returned by the server is applied, and a write that succeeded but stored something unusable now says so.
• Storage usage is more accurate. S3-backed files are counted, the nightly usage reconciliation is repaired, streamed downloads are no longer truncated, and a category that could not be measured is shown as such instead of as zero.
• A quota-refused upload explains itself, and a new workspace's storage quota follows its owner's plan.

Integrations

• 11 APIs were dropping a header their provider requires. Notion, LinkedIn Ads, Square, Xendit and several AWS services now send their mandatory headers, so calls stop failing while the credential looks active.
• Custom APIs can declare where their credential goes (any header, prefix, query parameter, basic auth or OAuth2) and send their declared static headers. Previously a bearer token could go out in the wrong header while registration reported success. A self-hosted integration's base URL can also carry its own scheme.
• Credentials for the SSH, SFTP and Database nodes. These nodes can now take their connection from a credential created on the credentials page.
• X accepts video in parts, and TikTok supports direct post with chunked upload. Large native X videos upload in chunks, and TikTok exposes the per-account posting options and says which scope a refused post needs.
• OAuth scope fixes. Figma no longer requests scopes the provider never grants, LinkedIn's two integrations share one OAuth product, Slack user scopes are requested correctly, and Composio points at its v3 API.
• Your own APIs are safe from catalog imports. A catalog seed never deletes a custom API, a seed sharing a name with one imports alongside it, and deleting a custom API no longer removes a secret it does not own.
• Renaming a credential is no longer refused because of an unrelated credential.

Security

• A database copy alone yields no secret and no replayable token. Credential fields are encrypted based on the shape of their name rather than a short fixed list, which now covers keys such as secret access keys, private keys and API secrets. Webhook, share-link and other capability tokens are stored encrypted and looked up by hash.
• Role checks fail closed. Four role-gate holes were closed where a missing or unknown workspace role was treated as a member with write access, including two reads that were reachable without an account on a self-hosted install.
• Tighter internal boundaries. The gateway strips attribution headers that services trust, the interface allow-list is enforced on every relay, CLI-bridge access is checked on every agent loop, and a catalog import can take a credential out of service but never delete it.

Self-hosting & install

• License change: back to the LiveContext Sustainable Use License 1.0. v0.2.0 through v0.2.15 were published under AGPL-3.0; from v0.3.0 the repository, images and package metadata use the Sustainable Use License. You can still use, modify and redistribute LiveContext CE at no cost, including in production for your own organization. Offering it to third parties as a hosted or embedded product that competes with LiveContext's paid versions is not permitted, and the copyleft obligation no longer applies. Releases already published under AGPL-3.0 keep that license. LiveContext is source-available; see LICENSE for the full text.
• Password reset by e-mail. Self-hosted accounts can now recover a forgotten password. The response never reveals whether an address has an account, and links are single-use and die on any password change. It needs SMTP: set MAIL_HOST and MAIL_PORT together (plus credentials and MAIL_FROM), as documented in docker/.env.ce.example. Without a relay no e-mail is sent.
• Interface and agent content restored. A storage mapping issue could make interfaces render as a blank page showing a number, and agent prompts read back as a number. This release restores the original content automatically.
• Installer and guides aligned. The npm launcher and the public export now use the same Compose and environment files, keep database settings and combined profiles, detect the actual frontend port, and document onboarding, server URLs, backups and updates. On Windows, install paths containing spaces now work when the CLI calls Docker.
• Image architecture documented. Release images are built for amd64; ARM64 support is documented as conditional.
• The agent CLI MCP server starts again, after a missing dependency was restored.
• Database migrations apply automatically on startup; no manual step is needed.

Fixes

• Mail node: a slow mailbox is no longer treated as unreachable, since connect and read timeouts are now separate.
• Lists and panels: a deleted resource disappears from every surface showing it, the side panel shows exact counts, and the notification bell stops showing disabled automations as about to run.
• Classify: a classification now records its answer, not only the question, and a node naming only a model is metered with the right provider.

Update an existing install

git pull
docker compose pull
docker compose up -d

The compose file pins this release's image tag, so run git pull first. Your data lives in Docker volumes and is kept across updates.

Fresh install

git clone https://github.com/livecontext-ai/livecontext-ce
cd livecontext-ce
docker compose up -d

Then open http://localhost:3000

Images

Pull from whichever registry you prefer, the digests are identical.

• ghcr.io/livecontext-ai/livecontext-ce:v0.3.0
• ghcr.io/livecontext-ai/livecontext-ce-frontend:v0.3.0
• ghcr.io/livecontext-ai/livecontext-ce-bridge:v0.3.0

• ghcr.io/livecontext-ai/livecontext-ce-screenshot-renderer:v0.3.0 (opt-in renderer profile)

• docker.io/livecontext/livecontext-ce:v0.3.0

• docker.io/livecontext/livecontext-ce-frontend:v0.3.0
• docker.io/livecontext/livecontext-ce-bridge:v0.3.0

• docker.io/livecontext/livecontext-ce-screenshot-renderer:v0.3.0 (opt-in renderer profile)

• quay.io/livecontext/livecontext-ce:v0.3.0

• quay.io/livecontext/livecontext-ce-frontend:v0.3.0
• quay.io/livecontext/livecontext-ce-bridge:v0.3.0
• quay.io/livecontext/livecontext-ce-screenshot-renderer:v0.3.0 (opt-in renderer profile)

Each image is also tagged v0.3, v0 and latest if you prefer to track a moving line rather than pin an exact release.