Ask an AI agent to fix a typo on your website and you’ll probably hand it an API key. Most APIs only offer broad keys, so that key can also delete the site. Then the agent reads a web page with a hidden instruction in it, and nothing stands between that instruction and your API.
Tokening puts something there. You issue a grant for one task, something like “this post, these fields, one use, fifteen minutes,” and the agent gets only the grant’s secret. Tokening keeps the real key in an encrypted vault, checks every call against the grant, and writes down what it decided and why. It’s one self-hosted Go program with a SQLite file, and it runs on Linux, macOS and Windows.
The Pitch in Four Numbers
- 0 API keys handed to the agent. It gets a grant secret for one task.
- 1 use per grant, spent before the call goes out, so two parallel calls can’t both get through.
- 8 refusal codes, each with a reason in plain words, logged for later.
- 1,000 parallel races on each of five platforms in the v0.1.4 release, with exactly one winner every time.
How It Works
You tell Tokening about a credential once, and it goes into the vault. When an agent needs to do something, you issue a grant scoped to that job: which target, which fields, how many uses, how long. The agent connects to Tokening, never to the real API. Each call is checked against the grant. If it fits, Tokening makes the call with the real key. If it doesn’t, the agent gets a refusal with a readable reason.
The one-use rule is the hard part, and it’s where the testing went. The use is spent before the call is sent. Fire two calls at the same moment and one wins, one is refused. The release tests ran that race a thousand times on each platform.
Every decision lands in an append-only, hash-chained log. Since version 0.1.2 that log works as a flight recorder. It starts the moment Tokening opens its database, with nothing to set up, and it’s there for audits, incident reviews and replaying what an agent actually did. The project is also open about the weak spot: someone with access to the database could try to edit the log, which is exactly what the hash chain is for spotting.
What It Works With
Version 0.2 forwards MCP calls over HTTP and stdio, and plain REST calls. It can sit in front of common stdio servers such as the filesystem and GitHub servers. Claude Desktop connects through tokening connect. The recorded demo runs against a throwaway WordPress in Docker: editing post 2 is allowed, and doing it again, editing post 3 or deleting post 2 are all refused, each with a reason. Two parallel calls on a one-use grant produce one winner.
Who It’s For
- Publishers and editors who want an agent to fix one article without the power to touch the rest of the site.
- Developers and small teams running agents against their own WordPress sites, internal services and SaaS accounts.
- Anyone worried about prompt injection, where a page the agent reads tries to steer it somewhere else.
- Teams that need an audit trail of what their agents did, and why each call was allowed or refused.
Limits and What’s Next
Two things are on the list and not out yet: grant requests that wait for a person to approve them, and ready-made templates for GitHub, Stripe and Slack. There’s no hosted version. Each release gets a source review inside the project, and the site says plainly that this is its own review and not an outside security audit. And no, there’s no cryptocurrency here. The name uses the older sense of a token, a claim check or ticket.
The Four-Question Screening
Every project featured here gets the same four questions, the way you’d screen a startup idea.
1. Where is the money in it?
Tokening is free under Apache 2.0, and the site lists no paid plan. The money is in where agents are heading. As companies let agents act on real systems, someone has to answer for what those agents can touch, and security and compliance teams pay for that. The natural paid layers are a hosted version for teams that won’t run their own, approval workflows and policy management for larger organisations, ready-made templates for popular APIs, and an independent security audit that buyers can rely on. Support contracts for regulated teams are another path.
2. What real problem does it solve?
Over-broad credentials. Agents need some access to be useful, and today the only access most APIs offer is all of it. One bad instruction, injected or not, and the agent can do far more damage than the job ever needed. Tokening shrinks the agent’s power down to the task and leaves a record of every call.
3. The technology in plain words
Think of a hotel valet. You don’t give the valet your house keys, your office keys and your car keys on one ring. You give them a valet key that starts the car and opens nothing else, and the hotel writes down when it was used. Tokening hands agents valet keys. The real key ring never leaves the safe.
4. Moat and differentiation
Identity companies and API gateways will move into agent access, and some already are. Tokening’s edge is that it’s small and specific: one binary you run yourself, grants scoped to a single task and a single use, races tested hard, and refusals an ordinary person can read. That suits small teams who won’t adopt a platform. The moat that can grow from here is the library of task templates for common APIs and a reputation for getting the hard cases right, starting with the one-use race.
Why It Matters
Agents are moving from writing text to taking actions, and permissions haven’t kept up. A grant per task, checked at the door and written in a log, is the kind of plain control that lets people say yes to agents without betting the whole account on them.
Follow the Project
Start with the quick start, watch the demo, get a build from the download page, or read the code in the Tokening repository on GitHub. Tokening is listed with the other open-source projects on SoftDevStudio.com, and more featured projects are in the Projects section.