Heimdall
The operator inside Zuse Bifrost: ask it about your projects, teams, webhooks and history in plain words, have it do the routine work, and confirm anything that changes what already exists. Beside every page, metered against the team's monthly AI budget.
Heimdall is an operator that works inside your workspace. You tell it what you want in your own words, "which projects are in development?", "pause the CRM sync endpoint", "what changed on my projects this week?", and it answers from your live record or does the work, through the same operations the pages, the REST API and the MCP server use. It runs as you, with your permissions, and anything that changes or removes what already exists waits for your confirmation.
Metered, never invoiced to you
Every answer is costed at the model's list price and summed against your team's monthly AI budget. The line under the composer says where the month stands, and at the cap Heimdall stops rather than running over. Usage & Budgets has the meter.
Where to find it
The panel
A sheet that slides in beside whatever page you are on and stays open while you move around. Press Ask Heimdall in the sidebar or Ctrl+/.
Inline assists
Ask Heimdall beside any long field of a planning record: draft it, improve it, shorten it, make it testable or rewrite it as Given, When, Then, then take the draft or drop it.
Quality review
Review now on a story, a requirement or a rule: Heimdall reads it the way a developer and a tester will, scores it out of 100 and lists what to fix, quoting the words and suggesting the rewrite.
Settings → AI
The month's standing, the statement per feature and model, the latest calls, and the cap and warning line the team owner sets.
Brief import
Paste a client's brief: Heimdall drafts epics with their stories, requirements, rules, terms, personas, stakeholders, risks and decisions, quotes the passage each came from, and you accept, edit, merge or reject each one from a review queue.
Research
Research the domain on a project's overview: Heimdall reads the web for what this kind of product needs, the rules that apply and the integrations the client will expect, writes notes with their sources and leaves suggestions you keep or dismiss.
The council
Convene the council on a story, a requirement, a rule or a brief import: Odin, Thor, Loki, Týr, Frigg, Baldur and Hel each read it through their own lens, Heimdall writes the verdict, and you accept or dismiss what they found.
Improvement sweep
Check the record on a project's overview, and every week on its own: Heimdall reads the record for a requirement without a measure, a story without a negative criterion, a rule without an example, an entity without an owner and a risk without a mitigation, and leaves a card per gap. No model, no cost.
Start a project in conversation
A paragraph about the client becomes the client, the project and its record as drafts, with the council reading the result in the thread.
What it does for you
Answers you would click five pages for
Which projects are live, who leads each, what changed this week, which webhook endpoints are failing, who is on the team and what their role allows. It calls the right reads and answers with the facts.
Routine work in one sentence
A webhook endpoint registered, a delivery retried, every notification marked read: made at once, with a button to open the page it touched.
A card before anything that changes
Pausing an endpoint, rotating its secret, deleting it, revoking a token: prepared, shown with its exact arguments, run when you confirm.
Proactive, on the house stack
When the record it reads is thin it says what is missing and offers to draft it, and every technical proposal assumes the stack Zuse Technologies builds on unless the brief rules it out.
Only what you can do
It holds exactly the permissions you hold. For a team member, Heimdall is bounded by their role, and every change is written to the activity log as Heimdall's, on your behalf.
The same tools as the MCP server
Every operation Heimdall runs is a tool on the MCP server too, so what it can reach grows with the catalogue, module by module.
The tool families land module by module, with the modules themselves: today Heimdall reaches your account, the project registry, the teams and their members, roles and groups, API tokens, webhook endpoints and deliveries, notifications and the activity log. When no tool covers what you ask, it says so and names the page that does.
A turn, step by step
You ask. Type in the composer, or tap one of the questions an empty thread offers. Write the way you would to a colleague who knows your projects.
It works in the open. Between your message and the answer a timeline grows as it happens: the model's thinking as a faded line you can unfold, every tool as a step with the arguments it used hanging off it and the time it took, and "Working for 3s · 2 steps" counting above. Once the answer is in, it folds to one line, "Worked for 4.1s · 3 steps", and opens with a tap.
The answer streams in, formatted as it arrives, with a button for anything it made or read that has a page, and two or three follow-up questions you can send with one tap.
A card when it matters. If the work changes or removes what exists, it stops first: a card names the action and its arguments and waits for Confirm or Cancel. Confirming runs it; cancelling costs nothing.
- A failing endpoint shows as dead-lettered deliveries.
A failing endpoint shows as dead-lettered deliveries. The delivery log filtered to dead answers it, and the endpoint list names each one.
- List webhook deliveries0.4s
- status · dead
- limit · 25
- List webhook endpoints0.3s
- limit · 25
One endpoint is failing. CRM sync has three dead-lettered deliveries since
Tuesday, all project.updated, each refused with a 503 from the receiving
host. Slack digest and Deploy board are delivering normally.
Nothing changed here: reads run at once and show no card. Tap Retry the dead deliveries and the next turn replays them, which adds to the record and runs at once too; tap Pause CRM sync and you get a card first.
Heimdall or MCP
Zuse Bifrost is agent-native: every operation Heimdall runs is also a tool on the MCP server, so Claude, ChatGPT, Cursor or an automation of your own can run your workspace too. Both roads lead to the same operations under the same permissions. They differ in who runs the model, where the conversation lives and what else is in the room.
| Heimdall | Your own agent over MCP | |
|---|---|---|
| Where it lives | Beside every page of the app | Wherever you already work: Claude, ChatGPT, Cursor, a pipeline, code of your own |
| Setup | None. Press Ask Heimdall | Connect once over OAuth from the assistant's connector settings, or mint a token |
| The model | Chosen and run by Zuse Bifrost, with a router that picks a lighter or a stronger model per message, or the tier you pick | Whichever your agent runs, with its own memory, files and other tools beside Zuse Bifrost |
| What it costs | Model usage at list price, against the team's monthly AI budget | Nothing from Zuse Bifrost; your assistant's own plan or API bill |
| What it reaches | Your projects, teams, webhooks, tokens and history, as you | The same, bounded by the token's abilities and project scope |
| Before a change | A card with the exact arguments; Review every change asks for new things too | Your agent's own approval rules; every tool is marked read-only, additive or destructive, so a good client asks before the destructive ones |
| The conversation | Kept in Zuse Bifrost for ninety days with every step, button and card | Kept by your agent |
| The record | Every change in the activity log as Heimdall's, on your behalf | Every change in the activity log as the token's |
The two are not exclusive. A team can press Ask Heimdall in the morning and run the same operations from Claude Code in the afternoon, and the activity log tells the two apart.
Where it is going
The operator is the first of Heimdall's surfaces. Each of the next ones arrives with the module it serves and with its own page here: inline assists on every editor, the brief import that drafts a project's planning scaffold from a paragraph, quality reviews on stories and requirements, the impact analysis of a change request, research into a project's domain, the opening page of a generated document, and, as their modules land, test generation, release notes, incident helpers, ticket triage and the Monday digest. All of them run through the same budget and the same ledger.
Questions
Related
- The panel — the sheet beside every page, end to end
- Usage & Budgets — what a turn costs and how the month is capped
- Permissions & Safety — who may use it and what the model is shown
- Settings → AI — the budget, the statement and the models
- MCP server — the same operations for an agent of your own
How is this guide?
Last updated on
Software Requirements Specification
What the SRS a project generates prints, section by section, where each section is read from, how the introduction is drafted, and how the team builds and tests against the final copy.
The Panel
Heimdall sheet beside every page, end to end: opening it, the context chip and the eye, the start screen, threads, the composer, the model tiers and the review switch, the work timeline and the streamed answer, buttons and follow-ups, confirmation cards and the budget line.