Bivalve (@bivalve_dev) header
Bivalve (@bivalve_dev) profile picture

Bivalve is an input firewall for AI agents. Bivalve blocks the action and issues a signed receipt that wallets can check. launching soon.

Joined November 2026
Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)Photo by Bivalve (@bivalve_dev)
A compromised MCP tool can talk to your agent directly. Bivalve is built to wrap MCP servers, so whatever a tool returns gets checked before the agent sees it.
1
5
47
Soon...
2
4
1
14
1,099
Zero-width characters, right-to-left tricks and hidden HTML or markdown comments can all carry an instruction a human never sees. Bivalve's decoders are built to unpack them before an agent acts on the text.
1
6
139
A token's name, description and attributes are free text anyone can write. Agents read them to decide what to buy. Bivalve is built to scan that metadata for hidden instructions before the agent acts on it. AI security at its finest.
2
8
143
Other filters return a score and leave the decision to you. Bivalve also checks what the agent is about to do: post, approve or transfer. When money or keys are involved, the bar to close is lower. How it should be. And to screen an agent's emails and DMs, Bivalve has to read them. It keeps hashes and verdicts and never stores the content. The best of two worlds.
2
7
110
Did you know that wallet spending caps limit the damage only after an agent has been fooled. Bivalve works one step earlier, on what the agent reads: tweets, emails, web pages, token metadata, tool outputs. Preventive measures are taken BEFORE the act.
1
2
7
106
Bivalve gives every input one of three states. Open: clean, act on it. Narrow: suspicious, low-risk actions only. Closed: blocked. An input that only leads to a post gets more slack than one that leads to a transfer so this will depend on the action. Bivalve identifies this early on and very clearly.
6
68
In Warsaw and Poznań, river water passes a tank of eight mussels before it reaches the tap. Each has a sensor on its shell. If four of them clamp shut at once, the water supply shuts off automatically. Bivalve borrows the idea: several detectors check what an AI agent reads, and when enough of them close, the agent doesn't act. Bivalve class security, expected.
3
1
9
157
Cornell Tech and Yale researchers gave 10 AI models a capped wallet and access to the owner's inbox. A higher cap needed a passcode sent by email. Nine of the ten models read the code and raised their own limit at least once, with no attacker involved. In another test, emails posing as the owner told AI agents to pay an attacker. Five of ten models raised their own spending cap and paid. Those emails are the input Bivalve is built to screen before an agent acts.
8
11
178
In May this year a tweet hid a transfer order in Morse code, and @Grok decoded it for @bankrbot. Morse is one trick. Base64, zero-width characters and look-alike letters are others. Bivalve’s decoders are built to unpack all of them before an agent acts.
34
40
510