Bindfort

BindfortGate AI Agent Tool Calls

Bindfort is a security and evidence gateway for the Model Context Protocol (MCP). Positioned between an AI agent and its MCP servers, it checks allow and deny policies before tools run, scans complete installed dependency trees for supply-chain risks, and creates HMAC-signed receipts after each decision. The product is aimed at teams that need stronger controls and verifiable audit trails around agent-driven automation. Bindfort offers a free MCP scan, while full product pricing has not been publicly disclosed. Its core policy and dependency-scanning features are marked as working, but runtime guardrails remain on the roadmap.

freemium
MCP securityAI agent securitydependency scanningtool call policyAI security gatewayMCP auditsupply chain securitysecure MCP serversMCP dependency scanning
Indexed
3.2 (0 Number of reviews)

Log in to rate the project

Try Now

AI agents are becoming useful precisely because they can call tools, access services, and carry out work on a developer's behalf. That also creates a difficult security question: can a team trust every MCP server connected to the agent? The Model Context Protocol has become an important way to link agents with external tools, but the surrounding ecosystem is still young and verification practices vary widely. Bindfort approaches that gap as an intermediary rather than another tool server. It sits between the agent and MCP infrastructure, examining requests before they reach an upstream server and preserving evidence about what happened afterward.

A gate before the tool call

Bindfort treats an agent's request as untrusted until it passes a policy check. Developers can define allow and deny rules, with deny decisions stopping a request before it reaches the MCP server that would have executed it. That distinction matters in production. A log entry created after a sensitive tool has already run may help with investigation, but it cannot undo the action. Pre-call enforcement puts the control at the point where a risky request can still be blocked.

This model is particularly relevant when an agent can reach several servers with different capabilities. A team might want to permit read-only operations while blocking commands that alter files, access a sensitive system, or invoke an unexpected tool. Bindfort's policy layer does not eliminate the need to review those permissions, but it gives the team a central place to express and enforce them instead of relying entirely on the agent's behavior or the individual server implementation.

The practical starting point is a conservative policy. Teams can begin with a narrow set of approved calls, observe which workflows are blocked, and expand access deliberately. A broad allowlist may be easier during a prototype, but it also weakens the value of putting a gateway in the middle. Bindfort is most useful when its rules reflect the real boundaries of the application rather than becoming a pass-through proxy.

The overlooked risk is often below the top-level package

Security review of an MCP server cannot stop at the name of the package being installed. Modern software pulls in layers of transitive dependencies, and a problem several levels down can still affect the application that loads it. Bindfort says it scans the complete installed dependency tree, rather than checking only the direct or top-level package. That is a meaningful focus for teams evaluating servers from unfamiliar sources.

The vendor's own research and demonstrations reported that every official MCP server it audited contained at least two high-severity advisories. That is a finding from Bindfort's sample and methodology, not a measurement of the entire MCP ecosystem, so it should be read as a warning rather than a universal statistic. The underlying lesson is easier to generalize: dependency risk can remain invisible when an evaluation looks only at the package a developer intentionally selected.

  • Pre-call policy checks can allow or block requests before an upstream MCP server receives them.
  • Deep dependency scanning examines installed transitive dependencies, not just direct package names.
  • HMAC-signed receipts provide a tamper-evident record of each policy decision.
  • A free MCP scan offers a low-commitment way to inspect a project before adopting the broader gateway.

Receipts turn activity into evidence

Many integrations produce ordinary logs, which are useful but not automatically trustworthy. Bindfort adds a signed receipt to each decision, recording whether the call was allowed or blocked and which rule matched. The receipt uses HMAC signing, and the vendor says its verification command, bindfort verify, can detect changes to the record. That makes the output more useful for audits, incident review, and debugging than an unstructured line in a general-purpose log file.

The distinction is subtle but important. A log says that an event was recorded; a verifiable receipt is intended to show that the recorded decision has not been altered since it was signed. It does not prove that the original policy was well designed, nor does it make a compromised runtime harmless. What it does provide is a clearer chain of evidence around the gateway's own decision process. For teams that need to explain why an agent was allowed to call a tool, that context can save time.

Bindfort's published performance note also deserves careful reading. The advertised 0.9-microsecond figure refers to a policy-path microbenchmark, not the full end-to-end latency of the gateway, network connection, MCP server, and tool itself. Real deployments should measure the complete path under their own workload. The distinction is a positive sign in the documentation, but it is not a substitute for production testing.

What is ready, and what still needs time

Bindfort publishes a readiness view that separates available capabilities from planned work. Deep dependency-tree scanning and built-in policy execution are marked as working, which covers the two most immediately useful controls: examining what is being installed and deciding whether an invocation can proceed. Runtime guardrails remain on the roadmap, however. That means the current product should not be presented as a complete runtime defense for every class of agent behavior.

That boundary matters for buyers. A gateway can restrict calls and preserve evidence, but teams still need sensible MCP server selection, credential management, network isolation, patching, and review of the permissions granted to the agent. Bindfort addresses a specific layer of the problem. It is not a replacement for application security or for understanding what a connected server can actually do once a permitted call reaches it.

External figures cited on the product site provide additional context, including claims from third-party scanning about command-injection exposure and large numbers of servers affected by a design-level MCP remote-code-execution issue. These figures are background evidence attributed to outside research, not Bindfort performance metrics. Readers should check the underlying reports, definitions, and scan dates before using them to compare products or estimate their own exposure.

For a team running an agent that connects to multiple MCP servers, a sensible evaluation path is straightforward:

  • Request the free MCP scan and review both direct and transitive dependency findings.
  • Start with default-deny or tightly scoped policies, then add exceptions only for documented workflows.
  • Test receipt verification and measure end-to-end latency in a staging environment before making the gateway part of a critical path.

Bindfort's current proposition is pragmatic: stop questionable calls before execution and leave behind evidence that can be checked later. The missing runtime guardrails and undisclosed full pricing mean teams should evaluate it as an evolving security layer, not a finished answer to MCP risk. For developers putting agents into real workflows, adding that layer early is still more responsible than discovering its value after the first unexpected tool call.

Pros & Cons

Pros

  • Blocks risky calls before they reach the MCP server
  • Scans transitive dependencies in the installed tree
  • Signed receipts support verifiable auditing
  • Free MCP scanning is available
  • Clearly distinguishes microbenchmark results from end-to-end latency

Cons

  • Full pricing is not publicly disclosed
  • Runtime guardrails are still on the roadmap
  • Some ecosystem risk figures rely on external scanning reports
  • Public deployment documentation remains limited

Frequently Asked Questions

What does Bindfort do?

Bindfort is a security and evidence gateway for MCP. It operates between an AI agent and MCP servers, checks tool requests against policies before execution, scans the installed dependency tree for risks, and creates signed receipts after decisions. The result is a central control point for blocking calls and reviewing why they were allowed or denied.

How does Bindfort help secure MCP tool calls?

It uses allow and deny rules to stop requests before they reach an upstream MCP server, rather than relying only on after-the-fact logging. Bindfort also checks transitive dependencies in the installed tree and generates HMAC-signed receipts for decisions. Those receipts can be verified for tampering, although teams still need broader controls such as credential management and network isolation.

Is Bindfort free?

Bindfort offers a free MCP scanning service that teams can request. The complete pricing for the product itself has not been publicly disclosed, so organizations interested in ongoing gateway or policy features should contact the vendor or monitor its official site for updated plans.

Who should use Bindfort?

Bindfort is aimed at development and security teams operating AI agents that connect to multiple MCP servers, especially where tool permissions, supply-chain risk, or auditability matter. It may also suit researchers evaluating MCP server dependencies. Teams with a simple local experiment may find the gateway unnecessary, while production deployments generally have more reason to centralize policy and evidence.

Which Bindfort features are currently available?

The published readiness information marks deep dependency-tree scanning and built-in policy execution as working. These capabilities cover dependency review and pre-call enforcement. Runtime guardrails are still listed on the roadmap, so buyers should confirm the current release and deployment details directly with Bindfort before depending on it for a particular production control.

Explore More

Similar Tools

AI-Native eCommerce Infrastructure

AI-Native eCommerce Infrastructure

AI-Native eCommerce Infrastructure is a StoreFrame-hosted control plane built for Magento and Mage-OS teams that want more than a locked-down hosting environment. It combines containerized services, live observability, and a browser-based Claude Code shell connected to real store data. Merchants can ask natural-language questions about products, customers, or revenue, while developers can investigate logs, metrics, and security signals from the same workspace. The service starts with a three-day trial and does not require a credit card. Its main caveats are equally important: the AI workflow depends on the customer’s own Claude subscription, and the platform’s policies and documentation are still developing.

Check

Check

Check is a preflight tool for AI coding agents that verifies commands before they run. It checks whether referenced packages, paths, functions, imports, and shell commands actually exist in the project or local environment, helping stop confident but fictional instructions before they create more errors. The tool works with Claude Code, Cursor, and Antigravity, although its installer is currently limited to Windows. Users receive 120 free checks per day, then pay per request at $0.0068 AUD with prepaid billing. Check does not upload repositories or source files, but it is closed-source and should not be treated as a security product. It catches false references, not legitimate commands that happen to be dangerous.

TrueCode

TrueCode is an innovative coding assessment platform designed for the AI era. Instead of banning AI, it integrates it into a full IDE environment where candidates tackle real debugging tasks. Its unique TruScore™ system evaluates not just the outcome, but also the candidate's judgment, verification, and AI interaction quality. With a rear-facing camera capturing desktop snapshots every 10 seconds, it generates transparent, explainable reports. Ideal for tech hiring, team evaluations, and educational settings, TrueCode offers free practice for candidates to build a verifiable skill profile.

AgentSite

AgentSite

AgentSite is a middleware designed for the AI search era, tackling the problem of single-page applications (SPAs) built with React or Vue being invisible to AI agents like ChatGPT, Claude, and Perplexity. It injects AI-friendly metadata such as meta tags, JSON-LD, and markdown mirrors into your site without requiring any code modifications. The service offers a free diagnostic tool and supports various deployment methods including Nginx, Express, and Edge environments.

Bodega One Code

Bodega One Code

Bodega One Code is a local-first AI coding IDE with a built-in chat and autonomous agents, supporting Ollama, OpenAI, Anthropic, and more. Personal use is permanently free, with parallel agents via Fleet, scheduled automation loops, air-gap mode, and full model freedom. If you care about keeping code on your machine, this one is worth a look.

Nadir

Nadir

Nadir introduces a verifier-gated LLM router designed to cut API costs without sacrificing quality. It routes requests to cheaper models first, then uses a calibrated verifier to score responses. If quality falls short, it escalates to a more powerful model. This approach claims up to 60% cost savings while maintaining 98% quality, offering an OpenAI-compatible, two-line integration for high-volume, varied complexity workloads.

Open-source Alternatives

guidellm: Open-Source Tool for Evaluating and Optimizing LLM Inference

guidellm is an open-source tool developed by the vLLM team to evaluate and optimize Large Language Model (LLM) inference performance in production environments. It offers stress testing, latency analysis, and throughput assessment to help developers identify bottlenecks and fine-tune deployment configurations. The project is primarily written in Python and licensed under Apache-2.0. At the time of collection, it had 1214 stars on GitHub.

ai-gateway: Unified AI Gateway Based on Envoy Gateway

ai-gateway is an open-source project built on Envoy Gateway, offering a unified API gateway to manage access to diverse generative AI services. It simplifies AI application integration and operations by providing features like load balancing, caching, and rate limiting for various AI providers. The project is written in Go and licensed under Apache-2.0.

go-micro: Go framework fusing AI agent harness with microservices

go-micro is an open-source Go framework that fuses an AI agent harness with microservices, supporting MCP, A2A, and multi-LLM integration. It is licensed under Apache-2.0 and primarily written in Go. As of the collection time, the project had 22,755 stars on GitHub.

Kun: Local-First AI Agent Workspace

Kun is a local-first AI agent workspace that unifies coding, writing, design, research, and automation through a shared GUI and TUI runtime. The project is primarily developed in TypeScript and has an 'Other' license. As of collection time, it has 4813 GitHub stars.

terax-ai: Lightweight Tauri-based Desktop Dev Environment

terax-ai is a Tauri-based desktop development environment with a size of only 7-8 MB. It integrates a GPU terminal, CodeMirror editor, Git tools, and multi-provider AI agents, offering an all-in-one development experience. The project is primarily written in TypeScript and licensed under Apache-2.0.

jar-analyzer: Open-Source GUI Tool for Java JAR Analysis with AI Assistant

jar-analyzer is an open-source GUI tool for Java JAR package analysis, featuring an integrated AI assistant. It offers robust capabilities like JAR DIFF, method call graph exploration, DFS call chain analysis, taint analysis, and control flow graph (CFG) program analysis. Ideal for Java developers and security researchers, it streamlines code auditing and reverse engineering tasks. The primary language is Java, licensed under GPL-3.0, with 2111 GitHub stars at the time of collection.