mcpField
B
ByMorning

ByMorning

Remote
bymorning.aiAI & Agents1Updated 1d ago

About ByMorning

ByMorning is a self-hosted AI gateway for teams connecting people, coding agents, and internal systems to approved models under workspace permissions, budgets, and auditing. It also exposes workspace-authorized tools to agents through OAuth.

Key features

  1. 01Self-hosted AI gateway
  2. 02Model routing and usage limits
  3. 03Workspace permissions and auditing
  4. 04OAuth access to workspace tools

Use cases

  1. 01Route coding-agent requests to approved models
  2. 02Give agents permission-scoped access to workspace tools

MCP support

ByMorning exposes its Tools through a remote MCP server at `/mcp`. Agents authenticate with OAuth as Workspace members and receive only the tools permitted to their member account.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://{bymorning_host}/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

Tools

Details

From the project README

ByMorning

Project status. ByMorning is provided as-is under the MIT License. The company that built it has closed, so there is no vendor, support, SLA, or maintenance commitment, and no one is obliged to review issues or publish releases. The project is community maintained: fork it, open pull requests, and run your own builds. Read SECURITY.md before reporting a vulnerability.

ByMorning is a self-hosted AI gateway. It connects your people, coding agents, and internal systems to the models you approve, under your permissions, budgets, and audit, inside your own cloud account. Documentation: https://bymorning.ai/docs.

Install

Every release image is signed keylessly by this repository's release workflow and published to ghcr.io/bymorning/bymorning, with signed evidence (SBOMs, scans, provenance) on the GitHub release.

WhereUse
One machine, to try itDocker Compose evaluation. Not for production.
AWS or AWS GovCloud (US)Terraform module, also on the Terraform Registry as bymorning/bymorning/aws.
Azure (AKS)Azure installer: Terraform and Kubernetes manifests.
Any Kubernetes clusterHelm chart, with your own database, object storage, and key.

Connect coding agents

  • Models. Point Claude Code, Codex, Cursor, VS Code Copilot, OpenCode, or Gemini CLI at your installation's OpenAI-, Anthropic-, and Google-compatible endpoints with a Service Account key. Spend limits, routing, and usage apply to every call. See Connect coding agents.
  • Tools. Add https://<your host>/mcp to the agent as a remote MCP server. The agent signs in with OAuth as a Workspace member and gets the Tools that member may use, under the Workspace's permissions. See Use ByMorning's Tools from an agent. server.json describes it for the MCP Registry as io.github.bymorning/bymorning.

Develop

To develop, with Bun and Node 22+, run from the repository root:

bun install --frozen-lockfile
bun run dev

Open http://localhost:3000; local sign-in continues straight to the workspace. The development host persists Postgres-compatible data with PGlite under .bymorning/local/pglite; Docker is not required for the database. dev runs the API watcher on port 3210 and Vite concurrently; the API runs migrations at startup. Build the execution image with bun run --cwd app/host setup when you need Code Interpreter. Local development requires no SST or AWS credentials; SST remains the supported AWS deployment tool, and the Azure installer deploys to AKS with Terraform.

To configure both application ports:

PORT=4001 WEB_PORT=4000 bun run dev

PORT controls the API; WEB_PORT controls Vite and the default login callback origin. bun run local is instead the bundled, one-origin development host. See the host guide for the deployment contract, the development defaults, Vite arguments, and database ownership.

Application bundles

One host program, app/host, serves every deployment. Its production build is what the root Dockerfile ships:

bun run --cwd app/host build
node app/host/dist/bymorning.mjs

It requires the deployment's storage and key selectors (STORAGE_PROVIDER, CIPHER_BACKEND), their companions, and the database, cookie, and origin settings documented in app/host/README.md. The development build adds PGlite, filesystem storage, the local cipher, and the Docker Code Interpreter and needs nothing set:

bun run --cwd app/host build:dev
node app/host/dist/bymorning.mjs

Distribute the complete dist/ directory, including the installation documents and the loadable plugin packages. The Docker image runs the production bundle with Node; no compiled application executable is built.

Database

Run all database operations from packages/core so Drizzle uses the Core configuration. sst shell supplies the selected stage's linked Database resource:

cd packages/core

bunx sst shell --stage <stage> -- bun db generate --name <migration_name>
bunx sst shell --stage <stage> -- bun db migrate
bunx sst shell --stage <stage> -- bun db studio
bunx sst shell --stage <stage> -- bun db:check

Generate a migration after changing the Core schema, then inspect the generated SQL and snapshot before applying it. Never edit, replace, or delete a migration that has been applied to any stage. Make follow-up changes with a new forward migration instead.

Apply migrations before deploying code that depends on them. SST deployments run the migration Lambda before updating the Server and stop the deployment when migration fails.

PostgreSQL migrations are forward-only and live in packages/core/src/database/migration/. The current development baseline is a destructive clean break; databases created from the previous migration history must be reset.

Terraform (Azure)

Azure installations deploy with Terraform and Kubernetes manifests. Follow the Azure installer: AKS, Key Vault, Blob storage, PostgreSQL, and a digest-pinned image mirrored into your registry.

Releases

A release is one Gateway image, delivered as a linux/amd64 OCI tarball pinned by digest, plus signed evidence (SBOMs, vulnerability and license reports, scans, SLSA provenance, SHA256SUMS). It is installer-neutral: every installer delivers the same tarball and verifies the same digest.

Release signatures are keyless. There is no signing key to store or rotate. The full-gate workflow signs the evidence with Sigstore using its own GitHub Actions identity: the signing certificate names the workflow run, and the signature is recorded in the public transparency log. A signature is a release signature only when that identity is full-gate.yml of this repository running on a v* tag (script/release/signing.ts); the installer export refuses anything else, and it verifies every signature with cosign before it packages a release. Anyone can check a release with the identity and issuer in release/signer.txt, and rebuild a release from its tag to compare digests: the build is reproducible.

Cutting a release:

  1. Set the version in the root package.json, app/host/package.json, server.json, and charts/bymorning/Chart.yaml (both version and appVersion, and the image in its Artifact Hub annotation). sh script/release/versions.sh and sh charts/bymorning/test.sh check they agree. Date its section in CHANGELOG.md; merge to dev.

  2. Tag that commit v<version> and push the tag. The tag must match the package.json version.

  3. Dispatch the signed build on the tag, then download its release artifact:

    git tag v<version> && git push origin v<version>
    gh workflow run full-gate.yml --ref v<version> -f sign=true
    gh run download <run id> -n release -D release
    

    The run builds both images, runs the image-security, image-boot, kind and browser suites, and signs the evidence. The artifact holds release/bymorning-<commit>.tar (the Gateway) and release/bymorning-code-interpreter-<commit>.tar (the code execution image; both gitignored), release/manifest.json, release/signer.txt, release/trusted_root.json (the Sigstore trust root, for offline verification), and the evidence in release/<version>/gateway/ and release/<version>/execution/. Each artifact's signing in the manifest records who signed: keyless (the identity and issuer), ephemeral (a throwaway key, every other run), or unsigned.

    On a v* tag the run's last job, publish, then pushes both images to ghcr.io/bymorning/bymorning and ghcr.io/bymorning/bymorning-code-interpreter under the version tag, checks each pushed digest against the manifest, signs them keylessly with the same identity, and attaches the SBOM to both and the provenance to the Gateway (script/release/publish.sh). A first push can create a package as private (1.3.2 came up public): check each one is public under the organization's Packages settings. If a package name was used before by another repository, delete the old package first; the new repository cannot write to it (permission_denied: write_package).

  4. Commit release/manifest.json, release/signer.txt, release/trusted_root.json and release/<version>/, and publish the GitHub Release:

    git add release && git commit -m "release: <version>"
    tar -czf bymorning-<version>-evidence.tar.gz release     # tarballs are gitignored; the archive keeps release/'s layout
    gh release create v<version> --notes-file <changelog excerpt> \
      release/bymorning-*<commit>.tar bymorning-<version>-evidence.tar.gz release/manifest.json release/signer.txt
    

    The two evidence folders hold files with the same names (SHA256SUMS, signer.txt, ...), and release assets must have unique names, so the evidence ships as one archive. The release notes of v1.3.1 show the verification steps.

  5. Mirror the AWS module to the Terraform Registry: run sh script/release/terraform-module.sh <directory> into a clone of bymorning/terraform-aws-bymorning, commit, and tag the same v<version>; the Registry picks up the tag. Do not run terraform init inside the export before committing (the script writes a .gitignore as a guard). Publish server.json to the MCP Registry: mcp-publisher validate, then mcp-publisher login github as a member of the bymorning organization and mcp-publisher publish.

  6. Render the security review package for the release and attach the PDF to the GitHub release. First update docs/security/review/release.json (its VERSION, date, test result, and scan summaries from the evidence):

    bunx playwright install chromium     # once
    bun run security:render -- --release release
    gh release upload v<version> build/security/<version>/bymorning-security-review-<version>.pdf
    
  7. Package the Azure installer from the manifest: bun installers/azure/export.ts --out <directory> (bymorning-azure-<version>.zip). It runs from the release build or a descendant that changes only release/, CHANGELOG.md, docs/security/evidence/, docs/security/review/release.json, and package versions (--force overrides), refuses evidence that was not signed by a release signer or whose signatures do not verify (--allow-unsigned exports a marked "UNSIGNED REHEARSAL"), and refuses to change the signer a target already pins (a different workflow or issuer) without --rotate-signer; a new release tag is not a change.

For your own builds, bun script/release.ts needs Docker with buildx and syft grype cosign gitleaks semgrep (brew install ...). It leaves the evidence unsigned unless you point cosign at a development key (BYMORNING_COSIGN_KEY=<cosign.key>, recorded as ephemeral); such a build is never deliverable.

The full-gate workflow (.github/workflows/full-gate.yml) runs the same build and evidence pipeline on every merge to dev and nightly with a throwaway key, then boots the tarball with the AWS and Azure environments, runs the Azure kind end-to-end, the browser suites, and the installer checks. Dispatched with sign=true on a v* tag it is the release build described above; on any other ref it signs keylessly but only to exercise the signing path, and the export will refuse the result. Every other run builds <version>-ci.<run> and records its evidence as ephemeral.

SST authentication configuration

Each deployed stage requires these values in SST; the repository does not provide secret defaults:

bunx sst secret set WEB_COOKIE_KEY --stage <stage>

Deployment-time OIDC settings (AUTH_ISSUER, AUTH_CLIENT_ID, AUTH_CLIENT_SECRET, AUTH_PROVIDER, AUTH_EMAIL_DOMAINS) are retired. Configure SSO at runtime instead: deploy, open Setup with the installation setup token — never logged; it is written owner-read-only to the server's setup-token file at startup (<data directory>/setup-token for the development host, /tmp/bymorning/setup-token in the image), or supplied by the operator as the SETUP_TOKEN secret — register the callback URL it shows with your IdP, then enter the issuer, client ID, client secret, and admitted email domains. Replace the connection later through Settings → Security (Auth Provider).

WEB_COOKIE_KEY must be a 32-byte base64url value. In commercial SST stages, the shared Router URL is the trusted application origin and serves the static frontend while forwarding /auth, /api, and /system to the Server. Terraform installations serve the embedded UI from the Gateway, which is the HTTPS origin. Local development uses its configured APP_ORIGIN; Vite forwards those same route families to the local API. Its local login provider and persisted cookie key are configured automatically by bun dev.

Any OIDC-compliant IdP with the authorization-code grant works. Each Auth Provider has its own redirect URI, ${APP_ORIGIN}/auth/callback/<providerID>, which Setup shows before you save. Enter the provider connection once in Setup; later changes go through Settings → Security (Auth Provider).

Integration OAuth callbacks use a process-local, one-use rendezvous. max: 1 prevents steady-state routing across multiple tasks, but it does not preserve pending flows across a restart or rolling replacement: every deployment or restart invalidates pending Integration OAuth attempts, and affected users must retry. Operations should drain active requests and avoid old/new task overlap; increasing the replica count requires shared callback storage.

Contributing

See CONTRIBUTING.md. Report vulnerabilities privately as described in SECURITY.md. Participation is governed by the Code of Conduct.

License

ByMorning is released under the MIT License. Third-party material is credited in THIRD_PARTY_NOTICES.md. The license covers the code; it does not grant rights to the ByMorning name or logo.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.