purserv0.1.0 · September 2026the purser for your coding agents
One local console for the coding agents you already pay for. Gated, recorded, and metered.
For anyone running more than one coding agent.
Clone and run
git clone https://github.com/purser-sh/purser
cd purser && bun install && bun run devA one-line installer ships with v0.1.0. What happens when you run it
9 agent adapters · 296 tests · Apache-2.0 · runs on your machine
What is this agent allowed to do, and what stops it?
Nothing stops it. Agents now run in the background, in parallel, and inside CI. Letting one run and checking the result afterwards is not a control.
Which agent changed this file, and who approved it?
The commit carries a human name. The model, the prompt, the tool call and the review disappear the moment the terminal closes. Nobody can prove what happened.
What did all of it cost, in money or in quota?
Every vendor reports only its own usage. On a subscription there is no dollar figure at all, only the quota you burned and the rest of the week to get through.
Someone edited a line out of the audit log. This is the CLI catching it, naming the entry and showing the hash that no longer matches.
Nobody can see what their coding agents actually cost. The prompt is the cheap part.
the agent loop
The prompt is a rounding error. Spend happens in the loop: file reads, tool results, context resent every turn, retries.
No vendor shows you the total across the agents you actually use.
~5,000
apps found leaking corporate or personal data. Without a record of what agents did, you cannot even ask which run caused it.
source
RedAccess scan of ~380,000 apps, May 2026
Every edit is staged as a diff before it is written. You approve it or you reject it, and there is no third path. The approval is enforced by the type system: the only function that can write to your workspace requires a token that only the approval step can create.
Budget caps are checked before every step, not after. Shell access is off by default, and when it is on, each command is classified against an allowlist, with anything unrecognised treated as dangerous and a git restore point taken before it runs.
Which agent, which model, which prompt, which file, approved by whom. An append-only ledger and a hash-chained audit log, so edits or truncation are detectable. Verifiable from the command line, in a file you own.
Every token across every agent in one ledger, with a real tokenizer per model. Metered API providers show dollars. Subscription agents show tokens, because your plan price is not knowable from the outside, and we would rather label that than invent a number.
We show dollars only where dollars are real. Metered API providers report in dollars. Subscription agents report tokens, because your plan price is not knowable from the outside. Every figure on your screen is labelled exact or approximate. We would rather tell you what we cannot know than show you a number we made up.
Not an editor.
Keep the one you have. Purser sits above it.
Not a model.
We sell none of them, which is exactly why we can meter all of them.
Not a cloud service.
It runs on your machine. No account, no telemetry, nothing uploaded.
Your provider's dashboard was built to show you its own spend, not everyone's.
Shows a diff before anything is written
Records which agent changed what, and who approved it
Stops a run before the bill or the quota is gone
Every agent in one number
Runs entirely on your machine
Price
No model vendor will ever show you Claude spend next to Codex spend next to Cursor spend. That is not a missing feature. It is a conflict of interest.
Free, forever.
One person, one machine. See what every agent costs you, cap them before they run away, approve every change before it lands, and keep a record you own.
Everything on this page is in the free version. Apache-2.0, no account, no telemetry.
git clone https://github.com/purser-sh/purser
cd purser && bun install && bun run devAfter v0.1.0.
One budget shared across a team. One audit trail for everyone. Policy set per repository, so a junior developer and a staff engineer do not have the same shell permissions.
Pricing is not decided yet. If your team has this problem, tell me and I will build it with you.
what it costs
Free and open source. Apache-2.0. Runs on your machine.
Purser doesn't resell tokens. You keep your own agent logins and your own bills. It just shows you what they add up to.
No model vendor will ever show you Claude spend next to Codex spend next to Cursor spend in one ledger. That comparison is against their interest.
purser ledger
append-only, hash-chained
Works with the agents you already have logins for. Purser meters them; it does not resell them.
Claude Code
Codex
Cursor CLI
Gemini CLI
Grok, Perplexity, OpenAI-compatible
Ollama
Echo
| Adapter | Kind | Auth | Metering |
|---|---|---|---|
| Claude Code | SDK / CLI | your login | tokens |
| Codex | CLI | your login | tokens |
| Cursor CLI | CLI | your login | tokens |
| Gemini CLI | CLI | your login | tokens |
| Grok, Perplexity, OpenAI-compatible | HTTP | your API key | metered, real cost |
| Ollama | HTTP, local | none | free |
| Echo | fake, for testing | none | not metered |
Subscription agents report tokens, not dollars. Your plan price is not knowable from the outside.
Seven documents in the product repo. Each link opens on GitHub.
Only what has shipped is listed as shipped. In a market full of demos that die on the second prompt, this section is the point.
17 shipped·6 in build·23 total