Latte 6be5ac3608
lint / lint (push) Successful in 39s
test / test (push) Successful in 37s
docker / lint (pull_request) Successful in 37s
docker / test (pull_request) Successful in 27s
docker / docker-test (pull_request) Successful in 11s
lint / lint (pull_request) Successful in 34s
test / test (pull_request) Successful in 32s
docker / docker-publish (pull_request) Has been skipped
Merge branch 'main' into feat/retarget-claude-mcp
2026-06-14 12:14:20 +00:00
2026-02-27 21:08:15 +01:00
.
2026-01-29 19:53:36 +01:00
.
2026-01-29 19:53:36 +01:00

AegisGitea-MCP

Security-first MCP server for self-hosted Gitea with per-user OAuth2/OIDC authentication for Claude, Claude Code, and Cowork.

AegisGitea-MCP exposes MCP tools over Streamable HTTP and a legacy SSE alias. Each user authenticates with Gitea through OAuth2/OIDC; repository authorization is checked per user before any service PAT call is allowed.

Securing MCP with Gitea OAuth

1) Create a Gitea OAuth2 application

  1. Open https://git.hiddenden.cafe/user/settings/applications (or admin application settings).
  2. Create an OAuth2 app.
  3. Set the redirect URI to this MCP server's callback: https://<your-mcp-domain>/oauth/callback.
  4. Save the app and keep:
    • Client ID
    • Client Secret

Required scopes:

  • read:repository
  • write:repository (only needed when using write tools)

2) Configure this MCP server

cp .env.example .env

Set OAuth-first values:

GITEA_URL=https://git.hiddenden.cafe
OAUTH_MODE=true
GITEA_OAUTH_CLIENT_ID=<your-client-id>
GITEA_OAUTH_CLIENT_SECRET=<your-client-secret>
PUBLIC_BASE_URL=https://<your-mcp-domain>
OAUTH_STATE_SECRET=<random-32-byte-minimum-secret>

GITEA_TOKEN is optional in OAuth mode. Without it, Gitea REST calls use the user's OAuth access token directly, so Gitea enforces permissions on every API call. With it, the token acts as a service PAT for API execution, but the MCP server first checks the requesting user's permission on the target repository through Gitea and denies the call if the user lacks the required read/write permission.

3) Configure Claude, Claude Code, or Cowork

Claude's hosted, desktop, mobile, Claude Code, and Cowork surfaces share the same remote MCP connector infrastructure. There is no Claude-specific server code path.

In claude.ai:

  1. Open Settings > Connectors.
  2. Choose Add custom connector.
  3. Paste https://<your-mcp-domain>/mcp.
  4. Complete the OAuth consent flow. Dynamic Client Registration (/register) handles Claude client registration.

In Claude Code:

claude mcp add --transport http aegis-gitea https://<your-mcp-domain>/mcp

Cowork uses the same connector model and MCP URL.

Manual OAuth client configuration remains available for clients that do not use DCR:

  • MCP server URL: https://<your-mcp-domain>/mcp
  • Authentication: OAuth
  • OAuth client ID: the client id returned by /register or your preconfigured client id
  • OAuth client secret: only for confidential clients

Hosted Claude callbacks are allowed by default: https://claude.ai/api/mcp/auth_callback and https://claude.com/api/mcp/auth_callback. Loopback redirects for Claude Code local development are allowed for http://127.0.0.1:* and http://localhost:*.

4) OAuth-protected MCP behavior

The server publishes protected-resource metadata:

  • GET /.well-known/oauth-protected-resource

Example response:

{
  "resource": "https://gitea-mcp.hiddenden.cafe",
  "authorization_servers": [
    "https://gitea-mcp.hiddenden.cafe",
    "https://git.hiddenden.cafe"
  ],
  "bearer_methods_supported": ["header"],
  "scopes_supported": ["read:repository", "write:repository"],
  "resource_documentation": "https://hiddenden.cafe/docs/mcp-gitea"
}

If a tool call is missing/invalid auth, MCP endpoints return 401 with:

WWW-Authenticate: Bearer resource_metadata="https://<mcp-host>/.well-known/oauth-protected-resource", scope="read:repository"

Architecture

Claude / Claude Code / Cowork
  -> Authorization Code Flow
  -> Gitea OAuth2/OIDC (issuer: https://git.hiddenden.cafe)
  -> Access token
   -> MCP Server (/mcp, /mcp/sse, /mcp/tool/call)
     -> OIDC discovery + JWKS cache
     -> Scope enforcement (read:repository / write:repository)
     -> Policy allow/deny
     -> If GITEA_TOKEN is set: check Gitea collaborator permission for <user, repo>
     -> Gitea API call with either the user token or the service PAT after authz

Example curl

Protected resource metadata:

curl -s https://<mcp-host>/.well-known/oauth-protected-resource | jq

Expected 401 challenge when missing token:

curl -i https://<mcp-host>/mcp/tool/call \
  -H "Content-Type: application/json" \
  -d '{"tool":"list_repositories","arguments":{}}'

Authenticated tool call:

curl -s https://<mcp-host>/mcp/tool/call \
  -H "Authorization: Bearer <user_access_token>" \
  -H "Content-Type: application/json" \
  -d '{"tool":"get_repository_info","arguments":{"owner":"acme","repo":"demo"}}'

Threat model

  • Shared bot tokens are dangerous:
    • one leaked token can expose all repositories reachable by that bot account.
    • blast radius is repository-wide and cross-user.
  • Token-in-URL is insecure:
    • URLs leak via logs, proxies, browser history, and referers.
    • bearer tokens must be sent in Authorization headers only.
  • Per-user OAuth reduces lateral access:
    • identity comes from Gitea OIDC/JWKS or userinfo validation.
    • without GITEA_TOKEN, API calls use the user's token and Gitea enforces permissions.
    • with GITEA_TOKEN, every repository-targeted call first checks the user's Gitea permission and fails closed if the check cannot be made.

CI/CD

Gitea workflows were added under .gitea/workflows/:

  • lint.yml: Ruff + formatting + mypy.
  • test.yml: lint + pytest + enforced coverage (>=80%).
  • docker.yml: lint+test gated Docker build, SHA tag, latest tag on main.

Docker hardening

docker/Dockerfile uses a multi-stage build, non-root runtime user, production env flags, minimal runtime dependencies, and a healthcheck.

Commands

  • make test
  • make lint
  • make format
  • make docker-build
  • make docker-up

Documentation

  • docs/api-reference.md
  • docs/security.md
  • docs/configuration.md
  • docs/deployment.md
  • docs/write-mode.md
S
Description
AegisGitea MCP is a private, security-first MCP (Model Context Protocol) server that enables controlled, auditable, read-only AI access to a self-hosted Gitea environment. The system allows ChatGPT (Business / Developer environment) to inspect repositories, code, commits, issues, and pull requests only through explicit MCP tool calls, while all access control is dynamically managed through a dedicated bot user inside Gitea itself.
Readme MIT 1.1 MiB
v0.2.0 Latest
2026-06-27 14:26:07 +00:00
Languages
Python 99%
Makefile 0.4%
Shell 0.3%
Dockerfile 0.3%