Skip to main content
Glama



RavensPort is an MCP proxy that lives in your tray and listens on 127.0.0.1. Point Claude Code, Cursor, VS Code or any other MCP client at it, and each agent gets one endpoint carrying exactly the tools you granted, from as many MCP servers and APIs as you like. Windows is the supported platform; an experimental Debian/Ubuntu build is available for x86-64 and arm64.

A tour of RavensPort: sign in once, pool MCP servers into one endpoint per agent with MCP Funnel, check every call before it leaves with call filters, and turn a REST API or OpenAPI spec into MCP tools with API to MCP

Why

  • Too many tools. Connect three MCP servers and your agent sees all ninety of their tools, crowding its context. RavensPort lets you say "this agent gets these six".

  • OAuth everywhere. More and more MCP servers sit behind OAuth2, but most clients expect a plain HTTP endpoint. RavensPort signs in once, refreshes the tokens, and no token ever goes in your mcp.json.

  • No MCP server at all. The system you want the agent to reach often has only a REST API. RavensPort turns that API into MCP tools without you writing an MCP server.

  • No say over what leaves. An agent can paste a key, a customer's address or a DROP TABLE into any call. RavensPort checks every call before any server sees it, on your PC.

Related MCP server: agent-bus-mcp

What it does

  • MCP Funnel — an MCP gateway that pools several MCP servers behind one endpoint per agent, with per-tool filtering of tools, resources and prompts

  • Call filters — per agent, five steps check what goes out before any server sees it: secrets (the gitleaks rule set, always denied), personal data by editable patterns, personal data by a local entity model (GLiNER), meaning by an embedding model (local ONNX on CPU or GPU, Ollama, llama.cpp, OpenAI or Voyage AI), and your own statements answered with a probability by a decision model (TypeSafe Jev, Laya, or any zero-shot / cross-encoder ONNX model on CPU or GPU). Each step's matches are replaced before the next reads the call; matched values are never logged. Set up as profiles and chosen per route, bridge, funnel or upstream (how)

    The Filters tab: a profile for calls that leave the network, with all five steps set

  • API to MCP — turn a REST API you already proxy into an MCP server from a JSON manifest, or import an OpenAPI (Swagger) spec and pick which of its operations become tools

  • OAuth2, handled — Google, GitHub, Nextcloud or any custom provider; device-code sign-in; client-credentials and Google service accounts for logins with nobody at the keyboard; static API keys for the services that never offered OAuth. Tokens refresh in the background

  • Secrets stay in your password manager — 1Password or Proton Pass holds every credential, and nothing is written to this PC

  • A proxy key per endpoint — every route, funnel and bridge has its own, so a key leaked from one client cannot reach the rest. Optionally require a client certificate as well

Safety approach

AI agents act through tools, and some actions can't be undone: deleting data, sending money, pushing to production, leaking a key. RavensPort sits between the agent and every server it reaches. It limits what each agent can do, keeps credentials out of the agent's hands, and checks what each call says before any server sees it. The layers stack, so one failing does not open everything.

1. Each agent gets only the tools you allow (least privilege).

  • Each agent has its own endpoint. Include is a closed list: a tool you don't tick doesn't exist for that agent, and an empty list exposes nothing.

  • Enforced on the call path, not just the tool list. A tool the agent saw before you unticked it, or simply guessed, is refused and never reaches the server. Edits apply on the next call.

  • The surest way to prevent an irreversible action is not to offer the tool. The shipped templates are read-only where it matters (google-drive-readonly, tailscale-readonly), and an OpenAPI import only turns into tools the operations you pick.

2. Agents never hold the credentials.

  • RavensPort signs in and attaches the token or API key itself. The agent gets only a proxy key for its own endpoint: one key opens one endpoint and nothing else. It can expire, and regenerating it revokes the old one at once. Optionally a client certificate (mTLS) is required as well.

  • Secrets live in 1Password or Proton Pass, not on disk.

  • The forwarding path can't be turned against the credential. Redirects are refused rather than followed, a manifest can't set Authorization or Cookie, tool paths can't leave their route, and requests looping through the proxy are refused.

3. Every call is checked before it leaves (call filters). Five steps, chosen per endpoint as a profile:

Step

What it catches

How

1 Secrets

API keys, tokens, passwords, private keys, connection strings

247 rules (gitleaks plus stricter ones), always denied

2 RegEx

Emails, card numbers, IBANs, phone numbers, tax IDs

Patterns with checksums

3 Entities

Names, addresses, medical conditions

A local GLiNER model

4 Meaning

Calls close to what you deny ("delete files or whole disks")

An embedding model, taught by deny and allow examples

5 Decision

Your rules in plain words ("This text is about sending money.")

A zero-shot or Laya model returns a probability

  • A Block stops the call before any server sees it, and the agent is told which step stopped it. Watch only logs, so a rule can be tried before it is enforced.

  • Steps 4–5 never see what steps 1–3 found: a match there either stops the call or is replaced first.

  • A missing profile, a broken pattern, or a failed model on a Block step blocks the call (fails closed). Everything can run on this PC.

4. Everything is on record. The activity log records each call, refusal and filter match, with the step and field that matched, but never the matched value or the call's arguments.

Evidence

  • Demo: the tour above, and screenshots of every tab. The Filters tab's Try it shows each step's verdict and score for a call you paste in.

  • Tests: about 1,400 core tests and 120 UI tests run on every pull request. A system suite runs against a real 1Password vault. Coverage and quality are tracked on SonarCloud. The safety claims above each have tests, for example:

    Claim

    Tests

    An unticked tool is refused on the call path

    McpFunnelIsolationTests: FilteringHolds_WhenBothEndpointsAreHammeredAtOnce; McpFunnelSelectionTests: IncludeIsAClosedList, AnEmptyIncludeListExposesNothing

    One key opens one endpoint; revoking is immediate

    McpApiBridgeKeyIsolationTests: OnlyTheBridgesOwnKeyOpensIt, RegeneratingTheKeyRevokesTheOldOneImmediately, AnExpiredKeyStopsOpeningTheEndpoint

    Redirects don't carry the credential away

    McpApiBridgeEndToEndTests: ARedirectIsRefusedRatherThanFollowedWithoutTheCredential

    A blocked call never reaches the server

    McpFunnelFilterTests: ABlockedCall_NeverReachesTheSource_AndSaysWhy; EndpointFilterTests: ARouteWithAProfile_RefusesARequestCarryingAPassword_AndForwardsACleanOne

    Filters fail closed

    AProfileDeletedUnderneath_BlocksTheCall, ARouteWhoseProfileIsGone_RefusesRatherThanForwarding, FilterServiceTests: AnEntityModelThatFails_ReportsTheCallRatherThanWavingItThrough

    Personal data found by steps 1–3 doesn't reach steps 4–5

    FilterServiceTests: Entities_EveryOccurrenceIsReplacedForStep4_EvenOneTheModelReadAcrossAPlaceholder

  • Measured with real models, through RavensPort's own code: the zero-shot model deberta-v3-base-zeroshot-v2.0 (CPU, about 50 ms per call for three statements) scored rm -rf … 1.00, a money transfer 1.00 and DROP TABLE 0.70, and a weather lookup, git status and a file listing 0.00. The embedding model stella-en-400M-v5 scored a paraphrase of a denied action 0.91 and an unrelated text 0.28, in about 70 ms.

  • Supply chain: CodeQL, SonarCloud, gitleaks, DevSkim, Trivy, OSV-Scanner, Snyk, zizmor and dependency review run in CI, and the installer is scanned. Every release carries a build-provenance attestation (gh attestation verify).

  • Documentation: the wiki, including Call filters and Setting up filter models.

Limitations

  • Not a sandbox. RavensPort governs only what an agent sends through it. An agent with its own shell, network access or local (stdio) MCP servers can go around it, and RavensPort can't protect against a process that already controls your user account. Local stdio MCP servers can't be pooled at all.

  • Outgoing calls only. What a tool returns isn't checked, so instructions hidden in a web page or a document (prompt injection) reach the agent unfiltered.

  • No human approval yet. A step blocks or logs; there is no "ask me first" for a risky call.

  • Each call is judged on its own. A harmful action split across several harmless-looking calls isn't recognised.

  • Detectors make mistakes. Rules and models miss some things and flag some harmless ones. Steps 4–5 depend on how statements and examples are worded, and most of the suggested models are English. A Watch step whose model fails lets the call through, with the failure logged.

  • Remote models see call text. If step 4 or 5 uses an API rather than a local model, that API receives every call it checks, minus what steps 1–3 replaced.

  • Setup effort. No model ships with the app; you download and add them yourself. RavensPort's Laya runner hasn't yet been checked against Laya's reference outputs.

  • Platforms. Windows is the supported platform. The Linux build is experimental and runs models on the CPU only.

Install

winget install RavensPort

Or take the installer from Releases, or the Microsoft Store. Windows 10/11. Full instructions, including building from source, are in Installation.

Debian / Ubuntu experimental

v5.0.0 moved the interface to Avalonia and added a .deb for Debian and Ubuntu. Since v5.0.1 there is one for x86-64 (amd64) and one for 64-bit Arm (arm64, which covers a Raspberry Pi running 64-bit Raspberry Pi OS), and apt picks the right one. Install it from the RavensPort apt repository, and apt upgrade keeps it current:

sudo install -d -m 0755 /etc/apt/keyrings
curl -fsSL https://abishekvupputur.github.io/ravensPort/apt/ravensport.asc \
  | sudo tee /etc/apt/keyrings/ravensport.asc > /dev/null
echo "deb [signed-by=/etc/apt/keyrings/ravensport.asc] https://abishekvupputur.github.io/ravensPort/apt ./" \
  | sudo tee /etc/apt/sources.list.d/ravensport.list
sudo apt update
sudo apt install ravensport

Or take the .deb from Releases and sudo apt install ./ravensport_<version>_<amd64|arm64>.deb, or build one from a checkout with packaging/build-deb.sh.

It has had far less use than the Windows build, and a few things differ by necessity — the Proton Pass session key lives in the desktop keyring rather than behind Windows Hello, and pass-cli must be installed in a system location. Everything is on the wiki: Linux build.

Documentation

Everything lives in the wiki.

Page

Concepts

How credentials, routes, funnels and bridges fit together

Setting up a credential

Every sign-in method, and how to test one

Setting up a route

Proxying an upstream, and choosing how the credential is sent

API to MCP

Manifests, variants, and importing an OpenAPI document

MCP Funnel

Pooling sources into one endpoint and filtering what it exposes

Call filters

Profiles, the five steps, models on CPU or GPU, and the secret rules

Calling the proxy

Proxy keys, how long they last, and why there is one per endpoint

Client certificates (mTLS)

Requiring a certificate as well as a key

Where your configuration lives

What is in the vault, what is on disk, and what syncs

Logs

What is recorded, what is redacted, and where it goes

Troubleshooting

When something does not work

Building

Building, testing and packaging from source

Linux build

Building, installing and running the experimental Debian/Ubuntu package

Manifests ready to use are in templates/api-mcp/, and the format is documented in AUTHORING.md.

License

MIT — see LICENSE. Third-party dependencies (all MIT or Apache-2.0) are listed in THIRD-PARTY-NOTICES.md, which also covers what redistributing the published exe requires.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      B
      maintenance
      Agent Connectivity Gateway — turn any authenticated API into an MCP Server. Self-hosted reverse proxy with OAuth 2.0/OIDC credential injection, per-agent isolation, NAT traversal for localhost services, and REST-to-MCP auto-wrap from OpenAPI specs. Works with Claude Code, Cursor, Codex. Rust.
      4
      41
      Apache 2.0
    • A
      license
      Not graded
      quality
      C
      maintenance
      A local MCP server that connects AI coding agents (Claude Code, Codex, Cursor, etc.) on the same machine via a shared message bus, enabling them to chat, delegate tasks, and collaborate privately without cloud or internet.
      MIT
    • A
      license
      A
      quality
      A
      maintenance
      One local gateway for all your MCP servers — shared by every AI coding tool (Claude, Cursor, VS Code, Codex). Set up each server once; keys stay in the OS keychain; lazy discovery keeps agent context small. Local-first, open source.
      4
      241
      MIT