self-hosting.md

Self-hosting

This guide is for operators evaluating Chopin on infrastructure they control. Chopin is an experimental research prototype, not a supported production service. Its current deployment model is one application process, one PostgreSQL database, one exact public origin, and a GitHub App registered for that origin.

Support boundaries

• GitHub.com is the only supported GitHub host.
• Chopin must be served at the root of one origin. Subpath hosting is not supported.
• HTTPS is required except for a loopback development origin.
• The application, API, MCP endpoint, and WebSocket share one origin and one internal port.
• One active Chopin process may write to a database. Horizontal application scaling and zero-downtime rolling deployment are not supported.
• PostgreSQL 17 is the version used by the repository and browser tests. Other versions are not covered by the project test suite.
• Database migrations are forward-only. There is no automated schema rollback.
• Restarting the application signs every browser out and releases hosted agent ownership. Documents, transcripts, decisions, background jobs and artifacts, research request staging, child documents, and implementation state remain. In local mode (see below), the returning browser can restore the same login without repeating device-flow sign-in; it still gets a new session and does not reclaim Planner ownership.

Choose the access policy first

Every browser user authenticates through the GitHub App. Two optional admission lists decide who may enter the instance:

• GITHUB_ALLOWED_USERS contains comma-separated GitHub logins.
• GITHUB_ALLOWED_ORGANIZATIONS contains comma-separated organizations whose active members are admitted.
• The lists are case-insensitive and form a union.
• Leaving both lists empty admits every verified GitHub user.

Admission does not grant repository access. Browser routes and WebSockets also require the repository to be present in a GitHub App installation available to the user. Pull access permits viewing; push or administration access permits creation, editing, and Planner invocation.

Chopin also registers /mcp unconditionally. It accepts a caller-supplied GitHub bearer token, applies the instance admission policy, and authorizes repositories directly from that token. It does not require a GitHub App installation and can create documents or advance implementation lifecycle state for callers with push or administration access. AGENT=off disables hosted agent turns and the background-job runner; it does not disable MCP.

Use restricted admission for any internet-facing evaluation unless unrestricted access is a deliberate choice. Protect the entire origin with TLS because MCP bearer tokens and browser sessions traverse it.

Choose and trust a harness

Chopin runs its hosted agent through @ai-sdk/harness. HARNESS selects one adapter from a code-owned map: the default is copilot-sdk, a host-process wrapper over @github/copilot-sdk; atomic embeds Atomic's headless SDK (@bastani/atomic) in the server process; and remote runs each Planner turn in a client attached over the agent connection instead. Adding an adapter to that map is a reviewed trust decision, not a runtime plugin choice: an adapter's builtinTools and supportsBuiltinToolFiltering are self-declarations, and Chopin's contract suite checks those declarations and the tools a session actually receives, but it cannot prove what the underlying runtime does internally. Only adapters that ship in this repository and pass that suite belong in harnesses.

For copilot-sdk, direct and ai-gateway are explicit HARNESS_AUTH modes that may be used on any bind. A mode that could fall back to a subscription already logged in on the host, such as auto, is refused at startup unless SERVER_HOST is loopback-only. The current copilot-sdk adapter still uses each owner's GitHub App user token and does not consume HARNESS_AUTH; setting auto on loopback passes the startup guard but does not enable host login in that adapter.

HARNESS=atomic runs Atomic 0.9.50 in the Chopin server process through its headless SDK (createAgentSession()). It does not spawn the atomic CLI or use RPC mode.

Choosing HARNESS=atomic gives the Planner shell and filesystem access as the server process's user, on hosted instances as well as local ones. Every Planner session is a full Atomic session: Atomic's workflows, subagents, MCP, web access, Intercom, and default coding tools run beside Chopin's document tools, with the operator's Atomic extensions, skills, prompt templates, and context files. There is no separate flag; choosing the harness is the choice. Operators who do not want that should use copilot-sdk, whose Planner keeps the isolated, Chopin-tools-only boundary. Only the summary and research workers still run isolated Atomic sessions. See The atomic Planner below and Hosted agent for the details.

HARNESS=atomic requires an explicit HARNESS_AUTH, either auto or ai-gateway. Unset, direct, and every other value are refused at startup.

• auto copies the host operator's Atomic login into memory when the first turn starts: $ATOMIC_CODING_AGENT_DIR/auth.json when that variable is set, then $PI_CODING_AGENT_DIR/auth.json for the legacy variable, otherwise ~/.atomic/agent/auth.json over the legacy ~/.pi/agent/auth.json. It also resolves provider keys from the process environment, such as ANTHROPIC_API_KEY, and ambient cloud credentials, such as an AWS profile or Google application default credentials. Because every admitted writer's turns would use the operator's own subscription or keys, auto is refused unless SERVER_HOST is loopback-only.
• ai-gateway starts without stored credentials, reads AI_GATEWAY_API_KEY, and accepts only vercel-ai-gateway models. It is the only mode allowed on a non-loopback bind.

The adapter never writes credentials to disk. It reads the host login without a lock file, keeps refreshed OAuth tokens in memory, and neither reads nor writes Atomic's models.json or models-store.json.

Under atomic, MODEL is required and must name a provider/model from Atomic's catalog: for example vercel-ai-gateway/anthropic/claude-sonnet-4.6 under ai-gateway, or github-copilot/gpt-6-luna under auto with a GitHub Copilot login. An integration that passes Atomic a failed model lookup gets another model without warning. This adapter looks the model up itself and fails the turn before any model request.

Structured output uses an inline Atomic extension that registers a terminating result tool, enabled only for a turn that requests structured output and blocked on every other turn. The tool records the calling model's own arguments. The adapter does not use Atomic's createStructuredOutputTool, which infers the answer again with a second model call. The Atomic contract suite (apps/server/src/harness/atomic.contract.test.ts) covers isolation, model resolution, host tools, structured output, and abort against Atomic's real agent loop. Run it before bumping @bastani/atomic.

Atomic caveats:

• Atomic's SDK defaults suit a local coding agent, not Chopin. By default it enables its shipped packages, including a mandatory Intercom, and its coding tools. It discovers host resources, gives turns without instructions a coding-agent system prompt, and saves tool results over 50,000 characters to a temporary file. For the isolated worker sessions the adapter overrides each default and fails closed on the ones it can observe, but a new Atomic release can add a default the suite does not check. Planner sessions deliberately keep those capabilities.
• Sessions live only in memory. doStop returns a state the adapter refuses to resume, so a session never survives a restart. Chopin does not resume harness sessions.
• Harness-level compaction, suspending a turn, and supplied harness skills are unsupported. A Planner session's own resource discovery does load Atomic skills. Atomic's automatic retries remain on.
• Since Atomic 0.9.42 a provider's rate limit is shared by the whole server process. After one request gets a 429, every Atomic session in the server, Planner turns and workers alike, waits for that provider's cooldown (its Retry-After, otherwise backoff of up to 60 seconds) before sending its next request, instead of each hitting the limit in turn. A turn waiting there shows no output yet; stopping it ends the wait without sending the request. Chopin sets no request budget (configureProviderRequestBudget()) or resourceAdmission of its own.
• Under auto, refreshing an OAuth token in memory can rotate the refresh token stored by the host CLI, which may then ask the operator to sign in again. auto also runs !command API-key entries in auth.json to resolve them.
• There is no per-session credit limit like Copilot's, so a worker's maxAiCredits does not apply. Use the provider's spend limits.
• @bastani/atomic adds more than 200 MB of installed dependencies on Linux x64, including glibc and musl native modules. Expect a larger image.
• Atomic workflows need no database server. Their runs are journaled in a SQLite file, $XDG_STATE_HOME/chopin/atomic/workflows.sqlite, beside the documents' Planner directories, which Chopin passes to each Planner session as its journal rather than through the process environment (see Full Atomic Planner). The Docker image sets XDG_STATE_HOME=/var/lib/chopin; mount a volume there to keep runs and those directories when the container is replaced.
• Local workflow worker pools stay off unless the operator's Atomic settings set workflows.localWorkers. Since Atomic 0.9.44 a pool runs as one coordinator process per root run, started by the Chopin server process, with the worker processes under it; it uses the same journal, and if the server exits, the coordinator and its workers exit with it. Atomic keeps each worker's standard error in worker-logs/ beside the journal, so under $XDG_STATE_HOME/chopin/atomic/worker-logs/<run id>/, inside the directory Chopin creates for the journal with access for the server's user only. It deletes a log when its worker shuts down normally but keeps the logs of workers that failed; remove them yourself. A child whose worker dies is replayed on a new worker, up to three attempts; after that the run fails and its card in Chat says so. The settings workflows.localWorkers.startupTimeoutMs and leaseMs set how long a worker may take to start and how long an unresponsive one keeps its child; Chopin sets neither. A pool runs children once one worker is ready and fails to start only when none is. Since Atomic 0.9.45 an unset startupTimeoutMs is no longer a flat 30 seconds: it is 30 seconds times the worker count plus the one-minute load average, divided by the CPU cores, at least 30 seconds and at most 10 minutes, so a large pool on a busy Chopin host waits longer. A run whose workers still miss it fails with Local worker startup timed out, and its card in Chat says so; raise the setting or use fewer workers.
• Since Atomic 0.9.44 a workflow runs up to 1000 of its stages at once unless it says otherwise (defaultConcurrency, 3 before), nests child workflows at most three levels deep (maxDepth, 4 before), and a stage's subagent tool runs up to 1000 children at once (parallel.concurrency and parallel.runningCapacity, 3 and 4 before). Without a pool, all of them run in the Chopin server process and share its memory and the provider rate limit above. To throttle a server, lower defaultConcurrency in extensions/workflow/config.json and the parallel limits in extensions/subagent/config.json under the Atomic agent directory, which Planner sessions read; Chopin sets none of these itself.
• Atomic depends on typebox 1.3.27 while Chopin pins 1.3.7. Host tool schemas pass to Atomic as plain JSON Schema, so the two versions never meet.

For a reviewed adapter that consumes a shared operator key, every admitted writer's turns would bill that key. Chopin adds no billing quotas of its own in this revision, across jobs or users; use the model provider's spend limits and usage alerts. A Copilot credit limit applies to one harness session, including all turns of its job stage, not as a platform-wide budget; see Background jobs.

The atomic Planner

A Planner session's working directory comes from one place: a checkout path supplied through the MCP invoke_planner tool. Chopin checks it with Git and compares origin's owner/repository to the document's repository, ignoring case and the remote host spelling (so SSH host aliases work). A mismatching or missing path refuses the invocation before anything is posted. A verified path is remembered for that document until the server process exits; nothing is written to storage. Every later Planner session for the document, whether started from the browser or through MCP, re-verifies the remembered path before using it. The check establishes repository coordinates, not remote-host authenticity, and it does not confine the shell.

Without a remembered checkout that still verifies, the session runs in a directory Chopin keeps for that document alone under its per-user state directory ($XDG_STATE_HOME/chopin/planner/<document id>, defaulting to ~/.local/state; ~/Library/Application Support/Chopin/planner on macOS; %LOCALAPPDATA%\Chopin\planner on Windows), with mode 0700. It is never shared with another document, keeps what the Planner and its workflows write there across later sessions and server restarts, and is not placed in the shared temporary directory. A symlink or file at that path is refused. The session has the same tools either way; the Planner is told whether it is in a checkout or in an empty directory without repository files, and its repository tools remain available.

Atomic input (ask_user_question, extension dialogs, and workflow-stage questions) appears as shared Decisions instead of terminal dialogs, including workflow run and stage labels. Unanchored batches append at the document's end. Dialog text and written answers stay verbatim, bounded only by the document's 256 KiB source limit and the shared-draft limits rather than the Planner's own question limits. A request has no time limit: it waits until a member answers it or it is withdrawn. Cancellation withdraws pending cards and never approves an action. See Full Atomic Planner for the mapping, resource loading, persistence, and workflow restart boundaries.

invoke_planner itself is offered on every harness and in both authentication modes. Other harnesses ignore its checkout: they neither verify nor use nor remember it.

The remote Planner

HARNESS=remote keeps the model out of the server. Each Planner session and turn goes to a client attached to the document over the agent connection at /api/v1/agent, usually an agent on a member's own machine; code, model credentials, and that client's tools stay on the member's machine, while the document, Decisions, and every room write stay in this server. The server needs no MODEL (one that is set is passed to the client as the turn's requested model) and accepts no HARNESS_AUTH, so it may bind to a public interface. Startup refuses HARNESS_AUTH. Summary and research workers run as isolated sessions on the client attached to their document, and wait while none is attached; set BACKGROUND_JOBS=off to keep them off the attached clients entirely, which also turns off /research. The route exists only with AGENT on and HARNESS=remote.

The trust boundary moves to whoever attaches. A client authenticates with a GitHub token bearer, checked like hosted MCP, and needs push or administration access to the document's repository, rechecked before every session and turn and every 5 minutes while it stays attached. The same account must have signed in to this server in a browser: attaching makes that login the document's Planner owner, one client per document, and a detached or dropped connection keeps it for PLANNER_ATTACH_GRACE_MS so the same account can reattach. The server pings every connection and counts one that answers nothing for AGENT_CONNECTION_TIMEOUT_MS as dropped, so a client whose network silently went away does not stay attached. A queue of frames draining toward the client puts that off, but never past AGENT_CONNECTION_MAX_SILENCE_MS since the client last sent anything, because a proxy such as Caddy can drain it while the client reads nothing. Its 15-minute default lets a client reading at about 19 KiB/s or faster drain the full 16 MiB the server may queue for it and answer a ping in time. A client that stops reading behind such a proxy is released by then at the latest, and usually much sooner: once the proxy's and the kernel's buffers fill, the queue stops shrinking and the ordinary timeout applies. Lower it to release such clients sooner, or raise it for slower connections. A frame counts only once it has arrived whole, so with the 2-minute default a client must upload at about 17 KiB/s or more to send a full 2 MiB frame; see Pings and raise the timeout for slower uplinks. The client then receives what a Planner turn sees, including repository reads made with that owner's token, and decides what the Planner writes during turns writers start; its run reports never start a turn. The server never runs the client's tools and gains no shell or filesystem access from this harness. Without an attached client, @chopin and invoke_planner are refused with planner-not-attached and instructions to attach one. See Remote Planner for the behaviour and Agent connection for the wire format. The connection shares the application's origin and port with /ws and /mcp, so a TLS-terminating proxy such as Caddy needs no extra listener; its default reverse_proxy already passes WebSocket upgrades with their Sec-WebSocket-Protocol and Authorization headers.

Prerequisites

• Docker for the application image, or Bun 1.4.4 for a source deployment.
• A reachable PostgreSQL database and credentials with schema migration access.
• A stable DNS name with TLS termination and WebSocket proxying.
• Outbound HTTPS access to GitHub and the selected harness's model provider (the hosted Copilot service for copilot-sdk).
• When the Planner is enabled (the default), or CONVERSATION_PLAN=on, a decisions classifier and its server-side key, separate from the Planner's harness credentials: OPENAI_API_KEY for OpenAI's Decisions API (api.openai.com) or JEV_API_KEY for TypeSafe's Jev (api.typesafe.ai). See Decisions.
• A GitHub App owned by the deployment.
• At least one user with repository push or administration access.
• For copilot-sdk, an active Copilot entitlement for each user who may own a hosted agent session. For atomic, model credentials for the chosen HARNESS_AUTH mode instead.

Register the GitHub App

Create one GitHub App per deployment. Register the exact public origin:

Homepage URL: <APP_ORIGIN>
Callback URL: <APP_ORIGIN>/auth/github/callback
Setup URL:    <APP_ORIGIN>/auth/github/setup

Then:

1. Enable expiring user authorization tokens.
2. Leave OAuth during installation disabled.
3. Disable webhooks.
4. Make the App installable on any account.
5. Leave device flow disabled, unless you use local device-flow sign-in.
6. Give it these repository permissions, which Authentication explains:

Contents:        Read and write
Pull requests:   Read and write
Workflows:       Read and write
Checks:          Read-only
Commit statuses: Read-only
Issues:          Read-only
Metadata:        Read-only (automatic)

Organization admission also needs one organization permission.

Runtime configuration

Store production values in the deployment's secret manager or an owner-readable environment file outside the source tree. Do not bake .env or credentials into the image.

Variable	Default	Meaning
STORAGE_DRIVER	postgres	Storage adapter. postgres is currently the only accepted value.
DATABASE_URL	required	postgres: or postgresql: connection URL. It is not printed by Chopin.
APP_ORIGIN	required	Exact public origin, without credentials, path, query, fragment, or trailing slash. HTTPS is required unless the host is loopback.
PREVIOUS_ORIGINS	empty	Comma-separated origins this deployment was served from before APP_ORIGIN, each validated like APP_ORIGIN. Paired machines, browser logins, Origin checks, and document links from those origins keep working; see Moving to a new origin.
GITHUB_APP_SLUG	required	Lowercase slug from the App's public URL.
GITHUB_APP_CLIENT_ID	required	OAuth client ID, not the numeric GitHub App ID.
GITHUB_APP_CLIENT_SECRET	required (hosted)	OAuth client secret used for user-token exchange and refresh. Unused when AUTH_MODE=local.
GITHUB_ALLOWED_USERS	empty	Comma-separated admitted GitHub logins.
GITHUB_ALLOWED_ORGANIZATIONS	empty	Comma-separated organizations whose active members are admitted.
SESSION_ENCRYPTION_KEY	required	Exactly 64 hexadecimal characters. Encrypts the hosted OAuth attempt cookie and derives a separate key for session credentials in PostgreSQL. Keep stable across releases and outside the database; replacing it invalidates hosted logins. Local mode requires the configured key but uses separate unpredictable, HttpOnly attempt and browser-binding cookies.
AUTH_MODE	hosted	Set local for loopback device-flow sign-in with persisted credentials (see Authentication). Any other value fails startup.
DEMO_PERSONAS	off	Set on to enable the server-owned local demo collaborators (Lily and Felix). Requires AUTH_MODE=local on a loopback bind; otherwise, or for any value besides on and off, startup fails.
CHOPIN_LOCAL_CREDENTIALS_DIR	platform default	Local mode only. Overrides the plaintext-fallback credential directory (default ~/.config/chopin on Linux, ~/Library/Application Support/Chopin on macOS, %APPDATA%\Chopin on Windows). Must resolve outside the repository and process working directory.
SERVER_HOST	127.0.0.1	Source-process bind address. The image sets 0.0.0.0, which local mode refuses. A HARNESS_AUTH mode that falls back to a host-logged-in subscription is refused unless this stays loopback-only.
PORT	8787	Source-process HTTP and WebSocket port. The supplied image and health check expect internal port 8787.
MODEL	gpt-6.1-sol	Model requested for hosted agent sessions. Required under HARNESS=atomic, where it must be provider/model from Atomic's catalog. Optional under HARNESS=remote, where the attached client chooses its model and a set value is only passed along as the turn's request.
HISTORY_MODEL	MODEL	Model for History's one-line summaries of editing sessions, written when someone opens History on a document with a Planner owner. A small, fast model suits it. Resolved like MODEL by the configured harness.
HARNESS	copilot-sdk	Adapter name selected from Chopin's harness map (copilot-sdk, atomic, or remote). An unknown name refuses at startup. remote runs Planner turns in a client attached at /api/v1/agent; see The remote Planner. atomic gives every Planner session shell and filesystem access as the server process's user, hosted instances included; see Choose and trust a harness.
PLANNER_ATTACH_GRACE_MS	120000	Under HARNESS=remote, how long a detached or dropped agent connection keeps the document's Planner ownership so the same account can reattach, in milliseconds, an integer between 0 and 3600000.
AGENT_CONNECTION_PING_MS	15000	Under HARNESS=remote, how often the server sends each agent connection a WebSocket protocol ping, in milliseconds, an integer between 1000 and 600000.
AGENT_CONNECTION_TIMEOUT_MS	120000	Under HARNESS=remote, how long an agent connection may send nothing, not even a pong, before the server closes it as dropped, in milliseconds, an integer between 1000 and 600000 and longer than AGENT_CONNECTION_PING_MS. A whole frame must arrive within it: 2 MiB in 120 seconds is about 17 KiB/s of client upload.
AGENT_CONNECTION_MAX_SILENCE_MS	900000	Under HARNESS=remote, how long an agent connection may send nothing at all, not even a pong, before the server closes it as dropped, however much of what the server queued for it drains meanwhile, in milliseconds, an integer between 1000 and 3600000 and at least AGENT_CONNECTION_TIMEOUT_MS. A draining queue defers the timeout only up to this, since a proxy can drain it while the client reads nothing. The 15-minute default lets a client reading at about 19 KiB/s drain the full 16 MiB the server may queue for it and answer a ping; lower it to release a client that stopped reading behind a proxy sooner, or raise it for slower links.
HARNESS_AUTH	unset	Auth mode forwarded to the selected adapter. For copilot-sdk, direct and ai-gateway are allowed on any bind and auto requires a loopback-only SERVER_HOST; the adapter does not otherwise consume it. For atomic, it is required and must be auto, which requires a loopback-only SERVER_HOST, or ai-gateway; every other value is refused. For remote, any value is refused.
HARNESS_EXTENSIONS	unset	Extension or package paths every atomic Planner session loads, as Atomic's --extension flag would, separated by the platform path delimiter (:, or ; on Windows). A package's extensions, skills, and workflows all register. Each path must be absolute and exist; any other harness refuses the variable at startup. Background workers never load them. The code runs in the server process as its user.
AGENT	on	Set exactly off to prevent hosted agent turns, disable the entire background-job runner, and avoid Copilot CLI startup.
BACKGROUND_JOBS	on	Set exactly off to disable background job scheduling. AGENT=off disables the entire runner. Under HARNESS=remote, workers run on the document's attached client.
WEB_RESEARCH	on	Set exactly off to disable new public-web research while retaining durable requests, artifacts, and other jobs.
RESEARCH_EVIDENCE_TIMEOUT_MS	600000	Deadline of one public-web research evidence attempt in milliseconds, an integer between 120000 and 1800000. Web search calls take up to about two minutes each; the attempt budgets at most four calls against what remains of this deadline. The job retries once after a failed attempt.
COPILOT_CLI_PATH	automatic	Advanced override for the Copilot CLI executable. Applies only to the copilot-sdk adapter.
CONVERSATION_PLAN	off	Set exactly on to enable experimental conversation-derived cards. Interpretation runs independently of AGENT; Planner jobs still obey the agent and job settings.
PLANNER_VISUALS	off	Set exactly on to enable the bounded visual handoff for foreground Planner turns. Requires AGENT=on and a decisions model.
DECISIONS_MODEL	from the key present	The classifier that answers every decision, as an exact Atomic provider/id. Unset, it is typesafe/jev-latest when JEV_API_KEY or TYPESAFE_API_KEY is set, otherwise openai-decisions/gpt-6-luna when OPENAI_API_KEY is set. See Decisions.
DECISIONS_TIMEOUT_MS	30000	Per-request decisions timeout in milliseconds, an integer between 100 and 60000.
OPENAI_API_KEY	unset	OpenAI API key for openai-decisions/gpt-6-luna. Server-only.
JEV_API_KEY	unset	TypeSafe API key for typesafe/jev-latest. Server-only.
CONNECTOR_VERSION	the bundled connector's	The chopin release this server recommends: chopin update installs it by default, and a refused connector is told to run chopin update <version> with it. Defaults to the version in apps/connector/package.json. A cli-v or v prefix is dropped; anything but a version like 0.2.0 fails startup.
CHOPIN_SETUP_AGENTS	unset	Setups the Machines page offers before the built-in coding agents, as a JSON array. See Setup agents. The supplied Compose file forwards it.
CHOPIN_BUILD_RESUME_MS	600000	How long after it starts the server waits for a paired machine to come back and carry on with a build it was running before a restart or deploy, in milliseconds, an integer between 1000 and 3600000. Until then the build stays running and its agent's calls are answered unavailable; after it, a build no machine took back ends as disconnected. A machine that registered again without its build has one minute, whatever this is. The connector stops an agent after 10 minutes without reaching Chopin, so a much longer window mostly keeps a build nobody will resume showing as running.

Automatic participation is part of the enabled Planner, with no separate feature flag. The shared Chat control pauses automatic replies per document. Group turn-taking asks the decisions model; solo replies bypass it.

See Background jobs and workers for the combined AGENT, BACKGROUND_JOBS, and WEB_RESEARCH behavior and recovery model.

Foreground Planner writes on every harness and MCP document imports share visual assessment before acceptance. The decisions model judges whether a faithful visual is possible and helpful, then selects its native type or a comparison table. The author must supply the selected visual beside its passage before the write succeeds. Lists and nested prose are included, and there is no diagram quota. Failed or uncertain assessments refuse the write. Prose imports require a decisions model even with AGENT=off.

PLANNER_VISUALS controls only the legacy single-paragraph tool for sessions without the shared admission check; it does not disable the shared check.

Conversation-derived cards default to off. When enabled, interpretation sends current and recent Chat messages and selected decision context to the decisions model, independently of the Planner's provider and credentials; AGENT=off does not disable it.

The same flag enables independent external-research suggestions in Chat. Their shared briefs can be synthesized by the research-brief@1 background worker under the existing Planner owner; this uses the configured harness provider, separately from the decisions model. A self-contained source excerpt is the fallback when synthesis is unavailable. Actual research starts only after a writer accepts the current shared brief and the execution capability is enabled. See Research offers from Chat.

The supplied Compose file forwards CONVERSATION_PLAN, PLANNER_VISUALS, DECISIONS_MODEL, OPENAI_API_KEY and JEV_API_KEY from the deployment environment. It uses the server's default 30-second decisions timeout. To override DECISIONS_TIMEOUT_MS, configure it explicitly in the service environment; the Compose file does not forward an unset optional timeout as an empty value.

Decisions

Conversation cards, research offers, group turn-taking, visual routing, and the living-build and prototype gates ask small typed questions: a yes-or-no probability, a choice from a fixed set, or a score on ordered levels. One classifier model answers all of them through Atomic's model registry. The server checks at startup that DECISIONS_MODEL is a classifier the registry knows and that its provider has a key, and refuses to start otherwise.

Any classifier in the registry works. Examples:

DECISIONS_MODEL	Key
openai-decisions/gpt-6-luna	OPENAI_API_KEY
typesafe/jev-latest	JEV_API_KEY or TYPESAFE_API_KEY
cloudflare-workers-ai/@cf/cloudflare/clef	CLOUDFLARE_API_KEY and CLOUDFLARE_ACCOUNT_ID

Keys come from the server environment under each provider's usual variable; see Atomic's model docs for the classifier list. The server never reads Atomic's saved logins. It does read Atomic's models.json from $ATOMIC_CODING_AGENT_DIR (default ~/.atomic/agent), so a provider's base URL can be overridden there. Requests send their decision context to the chosen provider and get one attempt each. CONVERSATION_INITIATIVE_MODEL can point group turn-taking at another classifier.

Generate the encryption key with:

openssl rand -hex 32

The image expects to listen on internal port 8787. Do not override PORT in the supplied image without also replacing its health check and container routing.

Setup agents

CHOPIN_SETUP_AGENTS adds your own setups to the Coding agent picker on the Machines page, before the built-in agents. Only signed-in people see them. Chopin refuses to start when the value is malformed.

[
	{
		"id": "acme",
		"name": "Acme",
		"install": ["curl -fsSL https://acme.example/install.sh | sh"],
		"command": "acme connect --server {origin} --key {key} --name {name}",
		"keyless": "acme connect --server {origin}",
		"default": true
	}
]

Field	Value
id	Up to 40 lowercase letters, digits and dashes. Unique, and not a built-in agent's id.
name	What the picker shows, up to 60 characters.
install	Up to 20 lines, shown under Show install steps.
command	The command for Cloud. It must contain {key}.
keyless	Optional. The command for Local, without {key}. Without it, only Cloud offers the entry.
default	Optional. true on one entry preselects it for anyone who hasn't picked an agent in that browser.

install, command and keyless also take an object with one value each for linux, macos and windows. The page fills in {origin}, {key} and {name}, quoted for the system's shell. With no machine name, it drops {name} and the flag before it.

The entry's command decides how the machine connects, so the page hides Run in the background for it. The list holds at most 10 entries.

Deploy the Docker image

The Dockerfile builds the browser client and one runtime image. The Bun server serves static assets, HTTP routes, /mcp, and /ws from the same process. Its default command applies migrations before starting the server and runs as the unprivileged bun user.

Build an image from a reviewed commit:

docker build --tag chopin:local .

Create an environment file such as /etc/chopin/chopin.env:

STORAGE_DRIVER=postgres
DATABASE_URL=postgresql://<user>:<password>@<database-host>:5432/<database>
APP_ORIGIN=https://chopin.example
GITHUB_APP_SLUG=<app-slug>
GITHUB_APP_CLIENT_ID=<client-id>
GITHUB_APP_CLIENT_SECRET=<client-secret>
GITHUB_ALLOWED_USERS=<comma-separated-logins>
GITHUB_ALLOWED_ORGANIZATIONS=
SESSION_ENCRYPTION_KEY=<64-hex-character-key>
AGENT=on
MODEL=gpt-6.1-sol
HARNESS=copilot-sdk
BACKGROUND_JOBS=on
WEB_RESEARCH=on

Restrict that file to the deployment account, then start the image behind a same-host reverse proxy:

docker run --detach \
  --name chopin \
  --restart unless-stopped \
  --env-file /etc/chopin/chopin.env \
  --publish 127.0.0.1:8787:8787 \
  chopin:local

If the reverse proxy is another container, attach both containers to a private network instead of publishing the application port. Ensure the database address in DATABASE_URL is reachable from the application container.

Reverse proxy requirements

The proxy must:

• terminate TLS for the exact APP_ORIGIN;
• forward the original Host and Origin headers;
• proxy WebSocket upgrades on /ws, and on /api/v1/agent under HARNESS=remote;
• proxy /mcp and the /api/v1/agent upgrade without removing their Authorization header, nor the upgrade's Sec-WebSocket-Protocol header;
• serve Chopin at /, not below a path prefix; and
• redirect alternate hosts to the canonical origin before application traffic, except the paths a moved deployment keeps serving.

Chopin derives OAuth callbacks from APP_ORIGIN, never from incoming Host or forwarded headers. A proxy cannot repair a mismatched configuration after the process starts.

Checked-in Compose files

The checked-in Compose files support repository development and the project's Coolify deployment; they are not a complete generic production stack.

compose.yaml publishes no host ports, uses a fixed internal development database credential, and includes Coolify late-binding variables. The local commands merge compose.local.yaml, which currently publishes application port 8787 and PostgreSQL port 5432 on every host interface:

bun run db:up      # start only PostgreSQL for source development
bun run docker:up  # build and start the application and PostgreSQL

Use those commands only on a trusted, firewalled development machine. Do not attach compose.local.yaml to an internet-facing deployment. Both bun run db:down and bun run docker:down tear down the whole Compose project.

Coolify supplies the public proxy and can substitute SERVICE_NAME_DB and SERVICE_FQDN_APP. Configure APP_ORIGIN as https://${SERVICE_FQDN_APP} and provide all GitHub, admission, model, and encryption values as runtime variables. Preview-specific late-binding and credential isolation are described in PR preview testing.

Deploy from source

A source deployment must build the client, apply migrations, and start the server as three distinct operations:

bun install --frozen-lockfile
bun run build
bun run migrate
exec bun apps/server/src/main.ts

All runtime configuration, including the GitHub App values, is required by the migration command. bun run start starts the server but does not build the client or migrate the database. A missing client build allows the API process to start while the browser route returns 404.

Run the direct server command under a process manager that restarts it after any unexpected exit. Some fatal runtime paths drain successfully and exit with code zero, so a policy equivalent to Restart=on-failure is insufficient.

Local device-flow sign-in

AUTH_MODE=local is a single-machine mode for evaluating Chopin without an operator-managed GitHub App client secret. It is not an alternative deployment topology: it requires a loopback SERVER_HOST and APP_ORIGIN, so it cannot be reached from another machine, and it is incompatible with the Docker image, which sets SERVER_HOST=0.0.0.0. Run it from a source checkout:

bun install --frozen-lockfile
bun run build
AUTH_MODE=local APP_ORIGIN=http://127.0.0.1:8787 GITHUB_APP_SLUG=<app-slug> \
  GITHUB_APP_CLIENT_ID=<client-id> SESSION_ENCRYPTION_KEY=<64-hex-character-key> \
  DATABASE_URL=<database-url> bun run migrate
AUTH_MODE=local APP_ORIGIN=http://127.0.0.1:8787 GITHUB_APP_SLUG=<app-slug> \
  GITHUB_APP_CLIENT_ID=<client-id> SESSION_ENCRYPTION_KEY=<64-hex-character-key> \
  DATABASE_URL=<database-url> exec bun apps/server/src/main.ts

GITHUB_APP_CLIENT_SECRET is not read in this mode. Enable device flow on the App instead of the client-secret authorization-code flow described above, and keep expiring user authorization tokens enabled. The browser shows GitHub's device code, persists the resulting credential to the OS credential store or, with explicit consent, a local plaintext file, and restores that login after a restart without a fresh device-flow prompt. See Local device-flow sign-in for the complete flow, the plaintext-consent warning text, file permissions, and the restart-restore and logout model.

Local logout deletes the persisted credential; it does not revoke the GitHub App authorization, because device-issued refresh tokens can be refreshed without the client secret but GitHub's revocation endpoint requires it. Remove the App from Settings > Applications on GitHub to revoke it there.

First-start smoke test

Startup validates configuration, database connectivity, migration history, the exclusive writer lease, and the configured harness (unknown HARNESS, or a host-login HARNESS_AUTH fallback on a non-loopback bind, refuse before the process serves traffic). It does not fully validate the GitHub App, Copilot entitlement, model, or lazy Planner runtime.

After the first deployment:

1. Confirm the process reports the intended restricted or unrestricted admission policy.
2. Complete GitHub sign-in and return to the exact configured origin.
3. Install the App on one non-sensitive test repository.
4. Confirm the picker lists only expected installations and repositories.
5. Create a channel with a user who has push or administration access.
6. Open the channel in a second browser and verify presence and live edits.
7. Send one @chopin request to verify model access (the owner's Copilot entitlement for copilot-sdk, or the HARNESS_AUTH credentials for atomic) and the hosted agent runtime.
8. Connect a local coding agent and call list_documents if MCP is part of the deployment's intended surface.

The image health check calls /api/session. It confirms HTTP liveness but does not prove that a browser bundle is present, a new database transaction can complete, GitHub is reachable, or the Planner can start.

Operations

Backups and restore

PostgreSQL is the durable system of record. Back up the complete database with the database provider's consistent backup mechanism; copying only selected tables or the application container is not sufficient. Regularly test restore into an isolated database before relying on the backup.

Stop Chopin before replacing a database from a backup. Starting against the restored database validates migration checksums and acquires a new writer lease. Browser sessions and Planner ownership are cleared at startup.

Upgrades

1. Back up the database and retain the currently deployed image.
2. Stop the existing application process.
3. Start the new image against the same database; its entrypoint applies pending migrations before serving.
4. Repeat the first-start checks after migrations or authentication changes.

Migration files are checksummed once applied. Never edit an applied migration. Because migrations are forward-only, returning to an older application image may require restoring the pre-upgrade database rather than merely changing the image tag.

Moving to a new origin

Paired machines' GitHub authorizations and browser logins are encrypted for the origin they were stored under, so changing APP_ORIGIN alone signs everyone out and makes every machine run chopin pair again. To move a deployment from https://old.example.com to https://new.example.com instead:

1. Add https://new.example.com/auth/github/callback to the GitHub App's callback URLs, keeping the old one, and point its Homepage and Setup URLs at the new origin.
2. Serve the new name, and keep serving the old one. On the old name, keep proxying /mcp, /api, /connector*, /auth/github/callback, and /auth/github/setup to Chopin, because paired machines, MCP clients, and GitHub still use it; redirect browsers on other paths to the new name.
3. Restart with APP_ORIGIN=https://new.example.com and PREVIOUS_ORIGINS=https://old.example.com.

A machine or browser login stored under a previous origin is re-encrypted under APP_ORIGIN the first time it is used, and the server logs once per kind that it moved one. Browser cookies belong to a host, so people sign in once on the new name. Origin and WebSocket checks accept the previous origins, and document links on them resolve to the same documents; Chopin itself only writes links on APP_ORIGIN. A sign-in that was in progress when the deployment restarted has to start again.

Remove PREVIOUS_ORIGINS once machines have reconnected and you stop serving the old name. Never list an origin you no longer control: its pages would pass Chopin's Origin checks.

Writer lease

The database holds one renewable chopin:writer lease. A second application process refuses startup. If the active process loses the lease, it drains and stops; fencing prevents an expired process from committing collaboration state. Allow the previous lease to expire before treating a failed host as safely replaced.

Troubleshooting

OAuth returns to an error page. Confirm APP_ORIGIN, the callback URL, and the browser origin match exactly. A trailing slash or reverse-proxy hostname change requires a configuration restart and corresponding GitHub App update.

No repositories appear. Authorization and installation are separate. Check that the App is installed on the repository, the signed-in user can access that installation, and any new permission is approved by the organization owner.

Organization admission is temporarily unavailable. Confirm the App has Members read access, the organization approved it, and the user has active membership. GitHub outages and rate limits fail closed for new checks.

A second process refuses startup. Another holder owns the database writer lease. Do not run two application instances against one database.

The process refuses startup with an unknown-harness or auth-mode error. HARNESS names an adapter that is not in Chopin's harness map, or HARNESS_AUTH falls back to a host-logged-in subscription while SERVER_HOST is not loopback-only. Fix HARNESS/HARNESS_AUTH or bind the process to loopback for local-only use. See Choose and trust a harness.

The UI returns 404 while APIs respond. The source deployment did not build apps/web/dist, or the runtime image was assembled incorrectly.

The first model-backed action fails. Check the invoking user's Copilot entitlement, the model, App permissions, repository write access, and Copilot CLI startup logs. Planner and research request execution validate these dependencies lazily.

MCP returns 401 or 403. A 401 indicates an invalid or expired bearer. A 403 indicates failed instance admission or a supplied Origin that differs from APP_ORIGIN. Repository authorization failures normally arrive as an MCP tool error with reason repository-forbidden. See Local agent MCP for client-specific guidance.

Related references

• Authentication
• Hosted agent (Planner)
• Local agent MCP
• Storage
• Repository channels
• exe.dev development