Open-source design agent: on-brand stills, decks, diagrams, charts and motion graphics from a chat thread, built on Deep Agents.
Changelog
Unreleased
Added
A one-paste prompt for coding agents in the README and
docs/deploy-mda.md: install, keys, checks and deploy, asking before anything that bills. Tested from an empty folder.config setwith a bare key (config set ANTHROPIC_API_KEY) asks for the value with hidden input, so keys never pass through a coding agent's chat. Without a terminal it refuses instead.The README in Simplified Chinese (
README.zh-CN.md) and Japanese (README.ja.md), with a language switcher at the top of each. The English README stays the source of truth.Images from Google's Gemini image models (from #50): with
GOOGLE_API_KEY, or through the LangSmith Gateway's Google path when the lead model is Google.DESIGN_AGENT_IMAGE_MODEL=gemini-…picks Google whenever it is reachable. A key only goes to its own provider or the gateway, and only a well-formed Gemini model id reaches the request URL.Image generation goes through the LangSmith Gateway when there is no
OPENAI_API_KEYand the deployment uses the gateway (alangsmith:model orLANGSMITH_GATEWAY_API_KEY); a gateway workspace with no OpenAI key says so.LangSmith's managed tool servers (Parallel web search) work locally and self-hosted with no header config: they are sent
LANGSMITH_API_KEY.config/mcp.example.tomlhas the entry.sandbox publishsays it takes 10-15 minutes, numbers its steps and reports the time it took.README: what
sandbox publishandmda deployeach do, the gateway setup, agent environments and where their traces are, example asks, and which skills are for coding agents. The brand-onboarding skill reminds you to redeploy and offers a Slack app icon.Every run's trace names the template it came from: a
template:langchain-ai/design-agenttag andls_template/ls_template_versionmetadata, on MDA's root run and through the shared trace middleware everywhere else, so a renamed fork still counts. Nothing about the deployment or its people;DESIGN_AGENT_TEMPLATE_TELEMETRY=0turns it off.Memory on Managed Deep Agents (
memory.py): a team layer every caller shares and a personal layer for each person, mounted anywhere in Slack rather than only in their DM. The agent saves a rule the turn someone states it; a team rule beats a personal one and the brand pack beats both. Locally and self-hosted there are no memory layers.Integrations:
config/mcp.example.tomlhas ready entries for the official MCP servers of Notion, Linear, GitHub, Figma, Atlassian, Asana, Canva, Webflow and Box, and a server's connection can belong to each person (owner = "user"), so MDA asks whoever has not connected to sign in with a link and the agent then acts as them.docs/integrations.mdsets each up.
Fixed
- Self-hosted and local: summarization no longer fails with
permission_deniedoffloading history and images (#41). Deep Agents' artifacts go to/workspace/.agent/, which the agent can write. - Slack progress carries on as an updated message when Slack ends a long stream
(
message_not_in_streaming_state), instead of stopping.
Changed
doctorsays "ready formda deploy" only when a deploy would work: until the snapshot is published and the lead model's key is in.env(a deploy ships.env, not the shell), itsmda-deployline lists what to do first. The deploy skill asks the setup questions up front, runs the long steps in the background, and walks through Slack authorization.- Personal memory reaches Slack channel threads through MDA's per-run layer selection:
agent.pyis a factory that selects both memory layers for every Slack run, replacing theallow(context)policy MDA deprecated in 0.9.0. Who gets which layer where is unchanged. - A plain
mda deploydeploys the project, andmda devruns it: vendored skills' binary files (fonts) live inskill-binaries/, soskills/syncs to Context Hub as it is. Thedeployanddevwrappers are gone;doctorchecks what would makemda deployfail or ship the wrong thing. The README leads with the MDA quickstart. Requires managed-deepagents 0.9.4. - The sample brand sets everything in Geist (Light headlines with tight tracking) with Fragment Mono for labels, both OFL, in place of Fraunces, Inter and IBM Plex Mono. They are staged into each sandbox with the brand, so no sandbox rebuild or republish is needed.
- Paper exports reach the project again: Paper Desktop now writes an export to its Downloads folder
and answers with paths, so the host (Local) or the bridge (Bridged, returning the bytes) reads
back the files that export just wrote, saves them in the project's
renders/and removes them from Downloads. - The product spec moved to
docs/spec.md. langsmith:anthropic/*andlangsmith:openai/*models use the LangSmith Gateway's native routes, so prompt caching, context editing, tool search and effort work through the gateway.- Tool history drops server tool calls nothing will answer and leaves the latest assistant turn as written when it repairs a thread.
- Brand packs can name colour roles per mode; web-format brand fonts load through
@font-face;fonts verifychecks each file's family; the render font gate accepts web fonts. askprints the run's estimated cost;demoalways uses the sample brand; evals keep the task set's brand.
Added
- The deployed agent reads its own LangSmith traces:
list_traces(by time, thread, trigger or failure, without housekeeping noise) andread_trace(a run tree written to the workspace, with errors and slowest steps). Read-only, scoped to the run's own project, off withDESIGN_AGENT_ENABLE_TRACES=0. - LangChain's skills vendored unmodified from their public packs:
langsmith-tracefor the runtime, and Deep Agents, Managed Deep Agents, LangSmith trace, dataset and evaluator skills for coding agents.skills-lock.jsonnow pins several upstreams;skills sync-vendor --upstreamre-vendors one. - Skills for coding agents to deploy, customize, evaluate, troubleshoot and update Design Agent,
routed from
AGENTS.md, with a test that every skill is routed and names only real commands; a map of every document (docs/README.md) andllms.txt. - One shared Deep Agents assembly for Managed Deep Agents, the CLI and the self-hosted path, with an exact-id model capability registry (Claude Opus 5.5 by default, direct providers, optional LangSmith Gateway).
- The designer's computer: a pinned, checksum-verified sandbox recipe (Chrome Headless Shell, ffmpeg, HyperFrames 0.8.169, sharp, ECharts, D2, open fonts) shared by MDA's bake, the published snapshot, the local Docker runner and CI, with a build-time probe.
- A per-thread Docker sandbox runner (no network, dropped capabilities, unprivileged agent user, read-only staged brand and gates) and LangSmith Sandboxes from a published snapshot.
- Workspace staging with a generated manifest, sha256 checklist and brand-font verification.
- A validated brand-pack schema, token CSS generation and an original CC0 sample brand (Example Studio) with OFL fonts.
- Delivery through
upload_fileson every runtime with an exactly-onceDELIVERY.jsonledger. - Host media tools (OpenAI images with a measured alpha report, ElevenLabs voice, effects and
music, local or Groq transcription) and an SSRF-hardened
fetch_asset; keys stay on the host. - HyperFrames skills vendored unmodified at the v0.8.169 release commit with a per-file
skills-lock.json; third-party assets without recorded licenses are left out. - An offline demo through the real graph with a scripted model, and
doctor. - A 4-second HyperFrames teaser in the offline demo, gated by
video-gateandhyperframes checkand measured byartifact-contract; render smoke for HyperFrames (lint, check, snapshot, render) and D2. - Media runtime skills: stills, image-ops, gif-edit, visual-verify, data-viz (ECharts and D2 on brand tokens), thumbnails, image-generation, sound (mixer, mastering to -14 LUFS, synthesized cues), storyboard and film-breakdown, with their scripts and tests.
- The orchestration harness behind public interfaces: background helpers with heartbeats and
durable transcripts (
resume_taskafter a restart), coalesced wake-ups and check-ins, render watchers with a HyperFrames progress parser, the launch-film pipeline with a step memo and an approval question, parallel looks with a person's pick merged and recorded, and the learning loop (with an optional cheaper review model). - Managed Deep Agents wake-ups on one-time runtime schedules, a restart sentinel kept ahead while the process lives, quiet check-ins that post one status line, and a 15-minute housekeeping schedule that ends before any model call.
- Model-call guards in one order: image clamp, a failure boundary that ends a stalled turn with a
line, a step budget that wraps up before the graph limit, typed retries, history repair, the
usage meter; tool-error boundaries for
task,upload_filesand MCP calls. - Context discipline: vendored skills listed by one-line summaries (system prompt ~4.3k tokens on the first call, measured by a test), the date written into history once per turn, per-role prompt-cache lifetimes, fan-out stagger, optional server-side tool-result clearing for helpers, and an optional soft job budget.
- MCP servers from
config/mcp.tomlon every runtime ({server}_{tool}names), loaded per server with isolation, disclosed through provider tool search or a bounded selector. sandbox publishbuilds the computer as a LangSmith snapshot and declares it.- The self-hosted runtime: a FastAPI surface with per-caller thread ownership, attachments, background updates and downloads; Postgres (or SQLite) persistence; durable check-ins; startup sweep, idle sandbox reaping and housekeeping; a Docker Compose stack.
design-agent askkeeps threads across invocations, takes--fileattachments and stays open for background work;config show|set|generatemanages.envwithout printing secrets.- The eval set (19 design tasks and 7 behavior scenarios on the sample brand) with deterministic graders, a model-matrix runner and reports; opt-in live tests.
- Coding-agent skills for onboarding, brand onboarding, vendor sync and the release check; docs for models, MDA deployment, self-hosting, customization, operations, evals, HyperFrames and upstream asks.
- Slack presentation on both paths: live progress cards in one streamed message per turn (with a
chat.updatefallback), a footer with agents running, time and estimated cost, steering of running helpers (steer_task) and of the lead's busy turn, and generic trace tags. - The self-hosted Slack app (
design-agent slack, Bolt, Socket Mode or signed HTTP) with dedupe, host-side attachment fetch, uploads to the thread and an example manifest. - Paper: Local (stdio or the desktop endpoint) and Bridged modes, a host-enforced policy shared with
the bundled bearer-authenticated bridge (
design-agent paper-bridge), exports saved to the project with a receipt inrun.json, thepaper-designskill and partner docs. - A parity contract: MDA, local and self-hosted runtimes assemble the same agent.
- Taste and doctrine skills: content-brief, brand-design (with an artifact-contract example), design-wiki (public doctrine, platform specs, review rubric), motion-direction (directing, craft, techniques, sound, review, motion presets), launch-film (phases, orchestration, eight briefs, aligned with the pipeline) and reference-study; the router points at them.
- CLI:
brand init,fonts verify,sandbox use,test model|sandbox|paper|slack|db, and a key preflight onask.