Arbiter
omry/arbiter
Firewall de capacidades para agentes IA com controle de acesso via políticas
Ver no GitHubServidor MCP que funciona como um firewall de capacidades entre agentes de IA e serviços, controlando acesso via políticas. Implementa integração com SMTP (envio de emails) e IMAP (leitura de pastas), com descoberta de capacidades e auditoria de operações.
Instalação
npx convez-ai omry/arbiter1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Controlar acesso de agentes IA a serviços de email (SMTP/IMAP)
- Implementar políticas de segurança para operações de IA em serviços críticos
- Auditoria e logging de ações de agentes em integrações de email
Veredito

Arbiter é uma solução madura e bem estruturada para segurança de agentes IA, com foco em controle granular de acesso via políticas. Ideal para ambientes que exigem governança rigorosa de integrações de IA com serviços críticos como email.
README
A tradução para o português deste README ainda está na fila.
Arbiter
| Description | |
|---|---|
| Project | |
| Packages | |
| Code quality | |
| Docs and support | |
| License |
Arbiter is a capability firewall between AI agents and services. Today it exposes policy-controlled access through MCP and a client CLI; additional interfaces may be added later. The current service surface covers sending mail over SMTP and reading IMAP folders through explicit account policies.
Project Status
Current implementation status:
- MCP server over stdio, SSE, or streamable HTTP via FastMCP
- capability discovery with SMTP and IMAP account and operation metadata
- SMTP submission with configured sender identity, TLS/auth settings, text/HTML bodies, and Bcc kept out of message headers
- IMAP list/get/search/move/mark-read/delete tools scoped to configured accounts and folders
arbiter.account.<service>and reusablearbiter.policy.<service>objects for per-service account policyarbiter-pyPython client CLI, nativearbiterclient CLI, andarbiter-serverserver/operator CLI- Docker deployment tooling for staged and installed Arbiter services
Known open gaps:
- durable audit storage is parked for post-v1, while startup/runtime logging is the v1 observability focus
- normalized error-code responses are still a design contract, while the implementation currently surfaces Python/MCP errors
- the agent-facing skill integration path is intentionally not implemented in this repository yet
Development
Create and use the repo-local virtualenv with:
python3 -m venv .venv.venv/bin/python -m pip install --upgrade pip.venv/bin/python -m pip install -r requirements-dev.txt
The repository root is a workspace, not an Arbiter runtime package. The dev
requirements file installs server, client/python-cli, plugins/smtp, and plugins/imap
editably so the arbiter-py and arbiter-server commands come from this
checkout.
Run the test suite from the repo root with:
.venv/bin/python -m nox -s tests.venv/bin/python -m nox -s lint
The lint session runs both black --check and pyrefly check.
For focused local runs without nox, use the same environment directly, for example:
.venv/bin/python -m pytest server/tests/unit/test_config.py.venv/bin/python -m pytest server/tests/unit/test_app.py
The Docusaurus website lives in website/:
cd website && npm installcd website && npm run startcd website && npm run build
The experimental native Go CLI client lives in client/go-cli/. Build all default release targets from the repo root with:
tools/build_go_client --clean
This writes stripped Linux, macOS, and Windows binaries for amd64 and arm64
under client/go-cli/dist/. Limit the matrix with one or more --target GOOS-GOARCH arguments, for example tools/build_go_client --target linux-arm64; pass --debug to keep debug symbols.
Package the native client as platform-specific skill artifacts with:
tools/package_arbiter_skill --clean
This writes a selector skill artifact plus one arbiter-skill-{os}-{arch}
target artifact per built binary under dist/arbiter-skill/, with local
directories and wheels generated side by side.
The native client can also be packaged as the arbiter-client PyPI project:
tools/build_release_dists --packages client
This builds one platform-tagged wheel per native target. Each wheel installs the
native arbiter executable directly as a wheel script; it does not contain a
Python wrapper.
The transitional Python CLI client lives in
client/python-cli/. It owns the repo-local arbiter-py
command for development, but it is not part of the release publishing surface.
The canonical distributable client is the native arbiter-client package.
The user-facing documentation lives in website/docs/. The root docs/ directory is reserved for internal planning and future design notes:
Local Streamable HTTP Run
For local Codex or VS Code integration, run Arbiter as a streamable HTTP MCP server and point the client at:
http://127.0.0.1:8025/mcp
Arbiter does not ship a runnable service config. Bootstrap a Hydra
config, edit it, then run the server. The default config directory is
~/.arbiter; pass --config-dir <dir> before a subcommand to use a different
location. config.local/ is ignored scratchspace for repository-local
development.
Plugin-owned object templates are created by the plugin command surface:
arbiter-server bootstrap arbiter
arbiter-server bootstrap plugin smtp account primary
${oc.env:...} reads the process environment that your shell, supervisor,
container runtime, or secret manager provides. For local runs, the root config
can name one optional dotenv-style file to load before composition.
See website/docs/operate/configuration-model.md for the generated file layout and composition flow.
For local development, a shell-owned env file can be useful:
# ~/.arbiter/local.env
SMTP_PRIMARY_ACCOUNT_USERNAME=agent@example.com
SMTP_PRIMARY_ACCOUNT_PASSWORD=change-me
ARBITER_IMAP_USERNAME=agent@example.com
ARBITER_IMAP_PASSWORD=change-me
Point the root config at the env file:
arbiter:
env_file: local.env
Existing process environment variables take precedence over values from the env file. Relative paths are resolved from the config directory.
Build or refresh that file from the active config:
arbiter-server env bootstrap
arbiter-server env check
env bootstrap keeps existing assignments, adds missing config references, and
groups entries under sorted # arbiter-* blocks plus # miscellaneous.
If arbiter.env_file is not set yet, it adds arbiter.env_file: .env to the
root config first.
Then run from this directory:
arbiter-server config check
arbiter-server serve
config check and serve require at least one configured service account.
Use arbiter-server plugins list to inspect installed service plugins before
validating a config. Once the server is running, use the client CLI against the
MCP endpoint:
arbiter-py mcp tools arbiter.mcp_url=http://127.0.0.1:8025/mcp
arbiter-py info arbiter.mcp_url=http://127.0.0.1:8025/mcp
arbiter-py info plugins arbiter.mcp_url=http://127.0.0.1:8025/mcp
The client can also read the endpoint from a small config file:
~/.arbiter/arbiter-client.yaml.
arbiter:
mcp_url: http://127.0.0.1:8025/mcp
Override config values with Hydra-style key=value arguments after the
command, or bootstrap the client config:
arbiter-py bootstrap client arbiter.mcp_url=http://127.0.0.1:8025/mcp
IMAP operations use folder-scoped UIDs returned by imap:list_messages and
imap:search_messages; pass those ids back to imap:get_message,
imap:move_message, imap:mark_message_read, or imap:delete_message with
the same account and folder.
License
Arbiter is distributed under the MIT License. See LICENSE.