LiveContext Files

What is new

The release where your agents reach you outside the app. Agents, approval steps and unattended runs can now ask their question in Telegram, Slack, Discord, WhatsApp or Microsoft Teams and take the answer from there, and alerts leave the app by email or on a channel. It also fixes how a {{...}} resolves in every node field, keeps long step outputs whole, and makes the cloud link a paid feature (read "Before you update").

Covers everything shipped since v0.3.1 (2026-09-22).

Corrected in v0.3.3: linking to LiveContext Cloud is open to every plan again, Free included (a completed cloud onboarding is required). Only the cloud relays, which spend cloud money, need a paid plan. v0.3.3 also restores the current models of every provider.

Before you update

• Fresh installs and updates work again: MinIO now comes from pgsty/minio. MinIO stopped serving its images: Docker Hub and Quay both refuse anonymous pulls of minio/minio and minio/mc, so a new install failed before it could start and docker compose pull failed on existing ones. Every compose file now uses pgsty/minio and pgsty/mc, Pigsty's maintained build of the same MinIO server (same data format, same commands). Your files are untouched: an existing storage volume is picked up as it is. Nothing to do beyond the usual update commands below.
• Linking a self-hosted install to LiveContext Cloud now requires a paid cloud plan. Every paid plan can link; the Free plan cannot. An install linked through a Free account keeps its link, its tokens and its settings, but the link is suspended: cloud models through the relay, catalog and model bundle updates, and the cloud web search stop until the account owner upgrades. Paying restores the link on the next heartbeat, with no re-link. The cloud link page of your settings shows a banner that says so. "Connect to Cloud" now walks a new account through the cloud onboarding (and pricing when needed) before it returns to your install.
• The model catalog now carries only models that are actually available: 119 instead of 410. The model bundle and the seed shipped in this image hold the cloud's active models (Anthropic, OpenAI, Google, DeepSeek, TypeSafe Jev, and the Claude Code, Codex and Gemini CLI bridges). On upgrade, the models it no longer carries are hidden from the pickers: all Mistral, xAI (Grok), Qwen, Moonshot (Kimi), Z.ai, MiniMax and Perplexity models, plus retired older versions of the others. Anything that referenced one of them (an agent, a workflow node) runs on its replacement or the platform default instead of failing. Models you added yourself and the edits you made to a model are kept. To keep using a provider that left the catalog with your own key, an admin can add its models back as custom models in Settings > AI Providers ("Add model"); custom models are never touched by a catalog update.
• Retired models stop running through the relay. A linked install that calls a model the cloud has retired gets it swapped for its replacement instead of a failing run.

Chat channels: your agents reach you where you already chat

• Five services. Telegram, Slack, Discord, WhatsApp and Microsoft Teams can be connected in the new Settings > Channels page. Each workspace keeps an address book of destinations and marks one as the default.
• A question travels to the person, and the answer comes back. An agent running on a schedule, a sub-agent, or any run nobody is watching used to either show a card on a screen nobody had open or decide alone. With a channel connected, its question arrives in the chat, you answer from there (in Teams, through a short decision page the message links to), and the run continues. A sub-agent's question reaches its caller the same way.
• Approvals from the chat. The User Approval node can send its request to any of the five services, and the press on "approve" or "reject" in the chat is what decides. A late approval authorizes that request only, never a standing rule, and the request deadline is enforced where the decision is taken.
• Pick the destination like a credential. Each agent and each approval chooses the workspace default or one specific destination. Sensitive actions an agent may take are set inside its channel card.
• Connect by hand, for free. "Set up manually" sits next to "Set up with the assistant": the same three steps as a form (account, destination, test message), with no assistant turn to pay for.
• Security: every inbound callback is authenticated (Slack signatures are checked over the exact bytes Slack sent, with a 5 minute skew bound), and a request is only answered from the chat its message went to.

Notifications

• Alerts leave the app, without the spam. For each workspace and topic (workflow failures, credits, account issues, tasks) you choose where alerts go: email, the default chat channel, both, or the bell only. Email alerts are always available on a self-hosted install; on the cloud they start with the Starter plan (credit alerts reach every account). A failing production workflow opens one incident and sends one alert, then one "recovered" and one "failing again", instead of one message per failed run.
• In your language and time zone. Every alert, reminder, recovery and daily summary is written in the recipient's locale (the six app languages), with times in their own zone.
• Inbox rows carry an icon each, with a red one for failed runs.

Workflows

• Every node field resolves {{...}} the same way. Fields that never resolved (filter and switch values, file names, the XML root element, sheet names, JWT payloads, classify and guardrail descriptions, response header values and more) now do. A templated number or boolean setting (a limit, a timeout, a count) no longer silently falls back to its default or resets the whole node config. Guardrail rules that were configured but ignored are enforced.
• A step keeps long text whole. API responses were cut to 200 characters plus "[TRUNCATED]" above 4 KB before being stored, so an email body or a scraped page reached the next node truncated. A workflow step now keeps text up to 1 MB.
• What the model received is visible. Classify and guardrail nodes no longer send their prompt twice, and the Params panel shows the prompt and content the model actually got.
• The builder keeps your layout. Node positions saved on the canvas are no longer overwritten by an agent building in parallel, interface edits show without a reload, and a plan reopens in the direction it was laid out in.
• Camera follow is on by default and now frames waiting nodes (interface, approval, wait) and follows an agent while it builds.
• Run logs open on all epochs and keep the id column.
• A long synchronous run (a video render, for example) is no longer reported as "could not start within 5 minutes" while it runs to completion.

Agents & chat

• Orbi, the general chat, has a face: the LiveContext mark perched on the composer, on desktop and on phones. It pulses while thinking.
• Orbi finishes with something you can see. Asked for a data workflow, it builds the table, the interface and a first run, then opens the result in the side panel with the new present action (application, run, workflow, table, interface, agent or file).
• Workspace member rules apply to the agent table tool. A read-only member can no longer write, rename or delete a table through an agent, and a member denied a table cannot reach it that way.
• An agent that sends an invalid update is no longer told it has hit its update limit.

Models

• Retire a model for good. An admin can retire a model; it then stays out of the catalog, the bundle, the seed and the relay until restored, whatever a feed sync or a migration does. A disabled model runs on the replacement set in the Models panel instead of failing every run that references it.
• The Models tab is virtualised and the rank can be typed.

Integrations

• Real logos. About 700 integration icons (letter tiles and outdated marks) are replaced by each brand's own artwork, with a dark variant where the default does not read on the dark theme.
• Scopes are checked before the call. 835 endpoints across 14 OAuth2 APIs declare the scopes they need, so a missing scope is reported before anything is sent or billed. A bring-your-own OAuth app can choose which scopes it requests, which unblocks Figma, TikTok and LinkedIn apps that do not hold every scope.
• Deleting your own OAuth app revokes only the connections it issued, not every account connected to that integration.
• A declared Accept header now reaches the provider, for every API in the catalog.
• Dead or misrouted endpoints fixed on Checkly, Affinity, Blueshift, Recurly and others; an unknown tool is reported as such, so an agent can act on it.

Self-hosting & install

• The API catalog update fits in 1 GB again. Since 2026-09-03 the catalog bundle (931 APIs, 32 MB) exceeded a JSON parser limit, so linked installs received no catalog update and re-downloaded it every 15 minutes. It now parses, and the apply streams one API at a time (peak around 460 MB instead of running out of memory).
• Your version is reported correctly. Released installs were counted as "dev" in the anonymous version ping; they now report their release number.
• Nothing that carries a credential (a token in a URL, a request body, a capability token) is written to a log line or returned in an error message any more.

Cloud accounts

• Two-factor authentication with an authenticator app, recovery codes, and a fresh sign-in before changing the second factor. A team owner is invited to turn it on.
• Sign in with SSO, routed by email domains a workspace has verified through DNS.
• One Free credit pool. The 1,000 monthly Free credits pay for workflow runs and for chat on the models opened to the Free tier; there is no separate AI allowance any more.
• Lower cost estimates on the pricing page. The figures are priced on DeepSeek Flash, whose price dropped (0.30 / 1.20 USD per million tokens, was 0.44 / 1.32), so a short exchange with an agent is now quoted at about 25 credits instead of 35, an agent conversation at 40 instead of 55, and a classify step at 1.6 instead of 2. Billing already followed the new price; the page now says so.
• A setup checklist after onboarding, a step to connect your apps during it, usage history in full screen, and richer usage analytics.
• A profile update can no longer change the account email or mark it verified.

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:v0.3.2
• ghcr.io/livecontext-ai/livecontext-ce-frontend:v0.3.2
• ghcr.io/livecontext-ai/livecontext-ce-bridge:v0.3.2

• ghcr.io/livecontext-ai/livecontext-ce-screenshot-renderer:v0.3.2 (opt-in renderer profile)

• docker.io/livecontext/livecontext-ce:v0.3.2

• docker.io/livecontext/livecontext-ce-frontend:v0.3.2
• docker.io/livecontext/livecontext-ce-bridge:v0.3.2

• docker.io/livecontext/livecontext-ce-screenshot-renderer:v0.3.2 (opt-in renderer profile)

• quay.io/livecontext/livecontext-ce:v0.3.2

• quay.io/livecontext/livecontext-ce-frontend:v0.3.2
• quay.io/livecontext/livecontext-ce-bridge:v0.3.2
• quay.io/livecontext/livecontext-ce-screenshot-renderer:v0.3.2 (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.