Clawvisor pricing, free plan and limits

Source-available gateway that vaults AI agent credentials and approves work by task. Free cloud tier, free self-hosting and a paid Pro plan for always-on agents.

Best for
Developers running agents such as Claude Code or Hermes who want a credential vault and per-task approval
Not for
CISO or compliance officers in regulated industries requiring SOC 2 or HIPAA certification

Should you use Clawvisor?

Clawvisor suits teams that want agents to reach Gmail, GitHub or Slack without ever holding an API key. The gateway source has 286 GitHub stars, runs free on your own host, and the README still warns it is experimental and unaudited, so treat it as one layer of defence.

Credential vaulting with task-scoped approval: agents reach the listed services without ever seeing a key, and a person approves each task once.

The founder previously built the identity-verification company Berbix and led trust and safety engineering at Airbnb, per the vendor's about page and the Y Combinator listing read on 15 November 2026. The gateway is source-available and self-hostable, so switching costs are low.

Also for

  • Developers building or deploying AI agents who need to give those agents real API access without credential exposure
  • Security engineers at companies running agentic workflows who need an audit trail and enforceable access policies
Show 2 more
  • Individual power users running personal AI agents across Gmail, GitHub, and Slack who want granular approval control
  • Platform teams evaluating MCP-compatible infrastructure for multi-agent orchestration

Also not for

  • Enterprise teams needing a vetted, audited security vendor with SLAs

Main trade-off

Experimental and unaudited by its own README, with no compliance certification listed, so regulated buyers cannot rely on it as their only control.

Clawvisor pricing

PlanMonthly priceWhat it includes
FreeFree1,000 protected calls a month, No credit card, Full gateway: verify, inject, log, Credential vault, Per-task containment and revocation, Risk scoring on every task
Self-HostedFreeFree under the Elastic License 2.0, No seat caps or call limits, Runs in your own network, Same purpose verification, risk scoring, vault and audit trail
Pro$120/mo20,000 calls a month included, $120 a month on annual billing, $150 month to month, Auto-reload keeps packs topped up, Extended log retention, Priority support, Everything in Free
Teams and tailoredCustomVolume quote for more than 20,000 calls a month, Tailored deployments, Org accounts on the managed cloud

Pricing page read on 15 November 2026. The hosted Free plan has no card requirement. Pro is billed at $120 a month on an annual plan or $150 month to month, with 20,000 calls a month. Packs cost $0.025 a call at list, with up to 30% off at the largest size. Self-hosting is free. Volume deployments are quoted by sales.

Before you commit

  • Calls are metered: past the monthly allowance agents stop until a pack is bought, at $0.025 a call at list.
  • Self-hosting means running the gateway on a separate machine from the agent, which adds infrastructure and upkeep that the listed price leaves out.

What Clawvisor does

Clawvisor is a source-available gateway between AI agents and the services they call, with 14 adapters from Gmail to Stripe. An agent declares a task, a person approves it once, and the gateway injects vaulted credentials at call time, so the agent itself never touches a raw key or token.

  • Credential vault

    API keys, OAuth tokens and other secrets sit encrypted in the vault and are attached inside the gateway at call time, so an agent's context and logs never contain one.

  • Purpose-based task approval

    An agent declares a purpose plus the tools it needs, a person approves that task once, and each request is checked against it until the grant is revoked at task end.

  • Risk assessment

    Each task is rated low, medium or high on reversibility, scope and tool sensitivity, and the gateway flags anomalous parameters, scope creep and prompt-injection attempts in tool arguments.

  • Audit trail

    Every request, purpose declaration, decision and credential injection is logged against the task and the approving person, and cloud plans add longer retention.

Show all 7 features
  • 14 service adapters

    Adapters cover mail, calendar, files, code hosting, chat, payments and SMS, with four more marked as coming on the integrations page.

  • MCP server with OAuth 2.1

    Hermes, Claude Desktop and other MCP clients sign in over OAuth 2.1, and the approval wait defaults to 240 seconds.

  • Self-hosting

    The same gateway runs on your own infrastructure through an install script or Docker, listening on port 25297, with no seat or call limits.

Works with

  • Gmail
  • Google Calendar
  • Google Drive
  • Google Contacts
  • GitHub
  • Slack
  • Notion
  • Linear
  • Stripe
  • Twilio
  • iMessage
  • Dropbox
  • Granola
  • Perplexity
  • Claude Code
  • Claude Desktop (MCP/OAuth 2.1)
  • Hermes
  • Telegram (approvals)

Missing integrations

  • SendGrid (coming)
  • Jira (coming)
  • Salesforce (coming)
  • Airtable (coming)
  • HubSpot
  • Zendesk
  • Confluence
  • AWS SDK

Most users expect these and they are not currently supported.

Capabilities

  • AI agent security
  • Authorization
  • Credential vaulting
  • MCP
  • Source available
  • API gateway
  • Agentic AI

Strengths and limits

Strengths

  • Agents never hold credentials, which removes the common failure where keys end up in agent memory, transcripts or logs.
  • Approving a task once avoids approval fatigue: a person reviews intent instead of confirming every action in a long run.
  • The agent-agnostic design works with anything that can make HTTP requests, including Claude Code and Hermes, with no SDK or code change.
  • Self-hosting is free and the source is readable, so a security team can inspect the gateway and keep every credential inside its own network.

Limits

  • By the vendor's own account it is experimental and unaudited, with a warning against making it the only safeguard on critical systems.
  • No SOC 2, ISO 27001 or HIPAA claim appears on the vendor's security page, which blocks buyers in regulated industries.
  • Calls are metered: when the monthly allowance runs out, agent calls stop until a pack is bought, which is a risk for always-on agents.
  • The source is available under the Elastic License 2.0, which is not an OSI-approved open-source licence, so check it against your own legal terms.

How Hokai rates it

  • Ease of use7.0 / 10
  • Features7.5 / 10
  • Value for money8.0 / 10
  • Support5.5 / 10

Clawvisor alternatives

ToolPrice fromFree planRuns onRating
Clawvisor This page$120/moYesWeb3.5
LikeClawLikeClaw sandboxes AI agent execution across many models at once; Clawvisor instead vaults credentials and gates access to real productivity tools like Gmail and GitHub.$5/moYesWeb--
HiddenLayerHiddenLayer's job is scanning ML models for backdoors and prompt injection; Clawvisor's job is stopping agents from misusing the API keys they already hold.QuoteNoWeb--
CyeraCyera classifies and protects sensitive data at rest across an enterprise; Clawvisor authorizes what an AI agent can actively do with that data in the moment.QuoteNoWeb3.3

Clawvisor facts

Deployment
Cloud · Self-hosted · On-premise
Minimum skill level
Intermediate
Price band
Free
Languages
English
Training
Documentation · GitHub README · Discord
Maker founded
2026
Total funding
Y Combinator Spring 2026 (amount undisclosed)
Training-data policy
Not applicable: Clawvisor is an authorization gateway, not a generative AI model. It does not train on user data.

Architecture and specs

Architecture
API gateway (Go 1.25+, React frontend, PostgreSQL or SQLite)

Data and privacy

Source available (Elastic License 2.0)

Questions about Clawvisor

What are Clawvisor's pricing plans in 2026?

Clawvisor Cloud has a Free plan with 1,000 protected calls a month, and a Pro plan at $120 a month on annual billing ($150 if you pay monthly) that includes 20,000 calls. Extra calls are sold in packs at $0.025 a call before discounts, topping out at $350 for 20,000 calls. Bigger or custom deployments go through a sales quote.

Is Clawvisor free to use?

Yes, two ways. The gateway source is public under the Elastic License 2.0, and the vendor says self-hosting has no seat caps or call limits. The hosted Free plan needs no credit card but stops calls once the 1,000 monthly allowance is used, until you buy a pack or upgrade, so an agent that acts all day will outgrow it.

What should you use instead of Clawvisor?

LikeClaw is the closer pick for teams that want sandboxed execution across many models rather than a credential-vaulting gateway. HiddenLayer suits teams whose main risk is a compromised model file, and Cyera fits enterprises classifying sensitive data at rest. Clawvisor's own compare pages weigh it against Arcade, HumanLayer, Portkey and Langfuse, which target adjacent agent-tooling jobs.

How do I connect an agent to Clawvisor?

On the cloud, sign up at app.clawvisor.com, connect the services an agent needs, and add the Clawvisor skill so the agent fetches data through the gateway instead of calling APIs itself. Self-hosting uses an install script with a setup wizard, and the vendor advises running it on a different machine from the agent.

Where does Clawvisor keep my credentials?

In an encrypted vault. When self-hosted they never leave your network. On the cloud they are stored encrypted and swapped in at call time, so neither the agent nor the model context touches them. Compared with a direct OAuth grant, the trade is one extra hop per request and reliance on the gateway being up.

About Clawvisor

Clawvisor is an authorization gateway for AI agents, built by Clawvisor, Inc. and backed by Y Combinator's Spring 2026 batch. Founder Eric Levine previously built the identity-verification company Berbix, which Socure acquired in 2023, and led the engineering trust and safety team at Airbnb. The product answers one problem: agents that hold API keys in .env files or broad OAuth grants can do far more than the job in front of them needs.

Read the full background

The mechanism is purpose-based approval. An agent such as Claude Code or Hermes declares a task, meaning a purpose plus the tools it expects to use. Clawvisor rates the blast radius as low, medium or high, and a named person approves or denies it. Requests inside the approved scope pass, anything outside it or on the blocked list waits for a human, and the grant is revoked when the task ends. Real credentials stay in an encrypted vault and are attached at call time, so they never enter the model's context. Approvals arrive through the dashboard, Telegram or a phone notification, and an MCP server at /mcp lets Claude Desktop and other clients connect.

The integrations page lists these adapters: Gmail, Google Calendar, Drive and Contacts, GitHub, iMessage, Slack, Notion, Linear, Stripe, Twilio, Dropbox, Granola and Perplexity, with SendGrid, Jira, Salesforce and Airtable marked as coming. The same page says anything else can be added to the vault as a key. The Notion and Perplexity adapters, for example, put Notion workspaces and Perplexity research behind the same approval step.

Where it does not fit: the vendor's README warns the software is experimental and unaudited, and the security page lists source-available code, self-hosting, a vault and an audit trail rather than any certification. Buyers who need SOC 2 or ISO 27001 evidence will not find it there. Calls are metered, so a busy agent on the Free plan stops at zero until someone buys a pack. Data-at-rest security is a different problem, which Cyera addresses (the Clawvisor vs Cyera guide compares them). LikeClaw covers agent sandboxing, while model scanning is the job of HiddenLayer. Clawvisor's own compare pages name Arcade, HumanLayer, Portkey and Langfuse as the products buyers weigh it against.

On 15 November 2026 we read the eval results the vendor publishes in its repository, last run on 15 March 2026 with Claude Haiku 4.5: 238 of 249 cases passed, 93.9% on intent verification, and 42 of 42 prompt-injection cases were flagged. The vendor ran these itself on its own cases, so read them as its figure, not an outside test.

Our call: choose Clawvisor if you run agents that need Gmail, GitHub or Slack access and you can accept an unaudited gateway as one layer among several, starting on the free self-hosted or cloud tier. Skip it if you need audited, certified controls with an SLA.

Sources and guides

More AI Tools on Hokai

  • Dynamo AIEnterprise AI security platform with real-time guardrails, pre-launch red-teaming and an agent and MCP risk scanner, sold on custom quotes.
  • ZeroGPTFree AI text detector with sentence highlighting, plus plagiarism, paraphrasing and humanizer tools; independent tests put AI detectors under 80% accuracy.
  • LynoteLynote is an AI humanizer and detector workspace with YouTube transcripts, notes and document translation, all drawing on one study-credit balance.
  • Coram AIAI-native physical security platform: plain-English video search, gun detection and access control layered onto any IP camera. 4.9/5 on G2 from 189 reviews. Per-camera license, quote only.
  • CybleAI-native cyber threat intelligence platform rated 4.8/5 on G2 from 151 reviews, built by Y Combinator-backed Cyble for enterprise SOC teams.
  • VerihubsIndonesian AI eKYC and deepfake detection platform reporting 99.95% face-recognition accuracy and NIST FRVT testing, used by BCA and Maybank.