CHANGELOG · 14th September 2026
Your agent can use the whole of Annsa.
Every action you can take in Annsa is open to a coding agent over MCP, and each reply names the next action.
Every action you can take in Annsa is now open to a coding agent over MCP, and each reply names the next action on that priority.
An agent working on its own needs more than a ranked list. It has to place the work, write the part of the spec it is qualified to write, record what it ruled out, and hand the job on with enough context that whoever picks it up does not start again.
The MCP surface covers the product: claim work, write a spec section, note what you found, verify a Done When line, place a priority on a bet, hand off, merge, regroup and ship. Every reply names the environment, the workspace and who is calling, and a request carrying another workspace's id is refused in words. A terminal names its seat once and every later call on that connection carries it, so the board shows which of your agents is holding what.
How to use it
A coding agent that can only read a list hands the work back. The MCP surface covers what a person can do, so the agent can take a priority from claimed to shipped and leave a record of how it got there.
- Connect your coding tool to Annsa, then ask it what needs you.
- Have it claim a priority, write the spec sections it is qualified to write and note what it found.
- Read the reply. It names the workspace it reached and the next action on that priority.
Also in this release
Ask finds the priority you meant
Ask matches on meaning rather than the words you happened to use, and asks what you mean by a number before it answers.
A priority opened cold says what was ruled out
Options set aside stay on the priority with the reason each was dropped, so whoever picks it up next can see what was already considered.
One number for what needs you
The board carries one count of what is waiting on a person, with priorities an agent proposed in their own inbox beside it.
Your board stays live while you work
Work done in another tab, by a teammate or by an agent shows on an open board without a reload.
Improvements
- Priorities and the board
- Shipped work keeps its rank, so one column reads as history and work in progress together. The evidence chip counts by kind: people, paying customers and findings. A board of any size lists every priority on it, with the count to match. You can rank a new bet first, and fetch a priority by its number.
- Specs
- Notes on a priority are dated and append-only, and Done When reads as Given, When and Then. Filing keeps everything you type, at any length. Two agents writing different sections of one spec keep both edits, and a person's edit survives the next refresh. Agents can read the direction document and propose changes to it.
- Agents and sessions
- Annsa tells the human owner, the agent holding the work, the credential and the terminal session apart. A handoff carries what is true now, the next exact action and the options already set aside. An agent proposes how serious a finding is, so a board with no customer signal still ranks.
- Security and access
- Team management, MCP tokens, webhook delivery and stored spec content each sit behind their own access checks. Every access path carries a test that exercises it as a second principal, and a ratchet counts those tests so the coverage grows with the surface.
- Connections
- Connecting a repository creates its webhook with no extra setup. A Google Sheets preview passes on what Google says, in Google's own words. A rate-limited request answers 429 with the headers that say when to retry, and each agent credential carries its own allowance.
- Review
- Review findings that survive a merge land on the board as rows in plain words, each naming the pull request it was found in.