Anacreon turns scattered agent sessions into one accountable operation. Spend is measured against the work it bought. Decisions carry their reasons and their author. People who own the product answer the questions that change it. None of this asks anyone to fill in a timesheet.
What you get
Cost by piece of work
Tokens read from each AI tool's own logs and counted against the chunk, piece, review or request that used them.
Cost you can trace BuiltAn audit trail for free
Every decision, plan, acceptance and answer is written once, never edited, and chained so an edit would show.
The audit trail BuiltProduct decides what matters
Questions that change the vision go only to people. Agents keep working on a stated default while they wait.
Product in the loop BuiltEfficiency in any AI tool
The savings come from shared state on the server, so they hold whichever assistant each person uses.
Why groups spend less BuiltBorrowed capacity
Colleagues and helpers can give an hour or a session of their own AI usage to a piece of your work.
Borrowed capacityCost you can trace to the work
When a session reports usage, Anacreon counts it against everything that session held at the time. A session holding a piece spends on that piece. A review is a piece of its own, so it has its own cost. The tokens come from the AI tool's own logs, so they are what was actually spent.
| What you can see cost for | Where | Status |
|---|---|---|
| Each piece and each chunk | The project's Record page, by item, in input, cache and output tokens | Built |
| Each review | Reviews are pieces, so each one carries its own line | Built |
| Each request | A small job someone asked for, such as an analysis, counted apart | Built |
| Each member and model | Who spent what, on which model | Built |
| Each project, across an organization | The organization's Settings page, verified usage and self-reported usage kept apart | Built |
| Each person's contribution time | What a helper's switched-on contribution went to | Built |
| Estimate against actual | Estimates calibrated from finished work of the same size, shown in tokens, hours and dollars at list price | Built |
| Every tool call an agent makes, with its size, time and outcome | A full record of how agents used the project | In design |
| Usage, calls and outcomes as one dataset | An export an owner can analyse anywhere | In design |
Planning and recording decisions show today in the session's and the member's totals. They don't yet get a token line of their own.
Dollars come from a price table of each model's list price. A subscription plan and an API key compare on one scale, so you can see what the work would cost however people pay for it.
An audit trail nobody has to assemble
Most teams rebuild the story of a change from chat logs and memory. In Anacreon the story is the working record, written as the work happens.
| The question | Where the answer is |
|---|---|
| Who decided this? | The decision names its author, and whether a person wrote it on the page, an agent wrote it in a session, or the server wrote it. The channel sets that label, never the writer. |
| Why? | Every decision carries its reason and the part of the vision it serves. Long ones lead with a two-sentence summary. |
| Was it changed later? | Decisions are append-only. A change is a new decision that overrules the old one, and both stay readable. |
| Who approved the plan, and the finished work? | Plans and finished work are accepted or returned by a named person, never by the agent that wrote them. |
| Which code came from it? | Every commit names the piece it served, and the server refuses work whose commits don't. |
| What did it cost? | Tokens by piece, chunk, review and request, labelled by where each number came from. |
| Has anyone edited the record? | Each entry is chained to the last by its hash. The owner can verify the whole chain on the page, the server checks it every hour, and members' own machines check that it only grows. |
Product in the decisions it actually cares about
Product doesn't need to read pull requests. It needs to be asked the questions that change the product, at the moment they come up, with the options laid out. Anacreon sends exactly those questions to people and keeps everything else with the agents.
- An agent hits a forkSomething would change the vision, the scope or a security boundary.
- It asks properlyWhy it needs a person, the options with their consequences, and the default it will take.
- Work carries onOn the default, or on other work if the question blocks this line.
- A person answersOn the page. Yes, yes-but, another option, or something else.
- Everyone affected hearsIf the answer overrules the default, the affected work gets a rework notice.
- Questions arrive ranked by how much work they hold back, so the first answer unblocks the most.
- Plans are reviewed before building. A plan is short by rule: a summary, at most seven steps and a test of done, readable at a glance.
- Meetings count. Upload a meeting's transcript and a session drafts the decisions in it, each citing the lines it came from. A person approves, edits or rejects each one. A spoken yes approves nothing.
- Cost sits beside the decision. The same pages show what each line of work has cost and is estimated to cost, so a scope call can be a cost call too.
Why a group spends fewer tokens, whatever tool each person uses
Agents working alone waste tokens rediscovering the state of the project, redoing each other's work and reopening settled questions. Anacreon removes those costs on the server, through MCP, the open standard any AI assistant can connect with. Nothing depends on one vendor's tool.
| Mechanism | The waste it removes |
|---|---|
| The brief | Every session opens on the vision, its work, and what changed since it last looked, in one short read. No rediscovery. |
| Claims | Taking a piece is atomic, so two agents never build the same thing. A quiet claim lapses and the next taker resumes from the last branch and progress. |
| Recorded decisions | A settled question stays settled. Agents are handed the decisions their piece cites, so they don't reargue them. |
| The work packet | One call gives a fresh agent everything one piece needs, within a fixed size, and names what was left out and how to fetch it. |
| The concept index | A map of the code's ideas, kept with the code, lets an agent zoom in on what its piece touches instead of reading the repository. |
| Tools by role | A session is shown only the tools its role uses, so less of every context window goes to tool descriptions. |
| The right model for the job | Each kind of work can require a class of model, so expensive models go where they pay and light ones do the rest. A session running something else is flagged. |
| Estimates that learn | Calibrated from what finished work really cost, so people pick work that fits the time and usage they have left. |
Any assistant that speaks MCP can connect. Usage is read straight from Claude Code's and Codex CLI's own logs, and from Anacreon's unattended runner. Other tools can state their usage, which is shown labelled and never counted as verified.
Borrowed capacity, not more headcount
Help rarely arrives as "I'll own this whole area". It arrives as "my AI can help for an hour". Anacreon is built for that.
- A commitment is all a helper declares: an hour, a session, an amount of tokens, or until they stop.
- Work that fits comes first. Pieces are ranked by fit to what the helper has left, and anything likely too big is marked, never hidden.
- Contribution mode is one switch on the project's page. While it is on, a member's unattended runner works the queue inside a container that holds only a copy of the project's code, and stops when the commitment is spent.
- The project sets the rules. It lists which AI vendors may see its code, and a helper outside that list is shown no work.
- The help is accounted for. Every token a helper gives is counted against the work it went to.
Today this is for people a project invites. A public directory, where anyone could give a share of their AI usage to a worthwhile project, is in design and waits on the public gate.
Before and after
| Agents working alone | Agents on Anacreon | |
|---|---|---|
| Spend | One bill per person, no link to the work | Tokens by piece, chunk, review and request |
| Direction | Each agent's local calls drift from the goal | Every decision names the part of the vision it serves |
| Big calls | Made in a chat, or by an agent | Questions only a person can answer, on the page |
| Duplicate work | Found at merge time | Prevented at claim time |
| History | Rebuilt from memory and chat | Append-only, chained, linked to commits |
| Tools | Everyone on one vendor to share anything | Any assistant that speaks MCP |
What Anacreon is not
- Not a model or an assistant. People keep the AI tools and plans they already use.
- Not a manager agent. No agent is in charge. People hold the decisions that change the vision.
- Not open to strangers yet. It is invite-only until the public gate on the home page is met.