Docs

antiloki is a free vibe graphing IDE. You describe the work, the Leader draws the team as a graph, and the loop runs it on the AI subscriptions you already have. These pages describe what the app does today, and say where something is not yet built.

1. Start here

antiloki works on a project folder. Open one, and the Cowork screen shows that project's agents, threads and sagas. Ask the Leader for what you want, and it drafts a team as a graph. You edit the graph, set its budget, and run it.

Run it on a project

  1. Open a project folder. The project needs to be a git repository for threads and worktrees to work.
  2. Sign in to at least one AI CLI (see your AI seats). Antiloki only offers the ones you are signed in to.
  3. Ask the Leader in the Cowork chat, or draw a team on the board yourself.
  4. Set the budget on the run, then start it. Watch the nodes change state as they run.
  5. Read the results: the artifacts the run produced, the saga it was recorded in, and the cost of every node.

How a run goes

A run goes through four stages, and you can see each one on the board:

StageWhat happens
AskThe Leader drafts the graph from your words, or you draw it. The Leader is the only terminal agent in the graph.
CheckThe validator refuses shapes that cannot run safely. Nothing starts until the graph passes. See what the validator refuses.
RunEach node runs on the seat it names, or the seat the tier picks. Forks run side by side. Each node records its agent, exact model, time and cost. A run stops when it reaches its budget.
JudgeQA and audit nodes judge the work. A human gate waits for your yes or no before anything leaves the project.

2. The Cowork screen

Cowork is where you work. The left side holds the chat with the Leader. The main area holds the panes, and each agent in a pane runs in a real terminal, not a simulation.

The Cowork window: the Leader chat on the left, a team graph running on the right
The Cowork window: the Leader chat, and a team graph running on the right. Example data, rendered from the site's own design. Not a capture of a live project.

Panes and the Studio

Panes sit side by side in the main area. You can fold a pane to a thin strip to keep it without losing its place. A pane can hold a terminal, a frame, a diff, a preview, or one of the Studio tools.

The Studio, on the right, opens the other tools of the app. Some of them, and what each shows:

ToolWhat it shows
TerminalA real terminal, for an agent or for you.
Diff and reviewWhat changed, for a thread or a run.
FramesMockups made by agents, as editable frames. The content is cleaned first: scripts, event handlers, iframes and network addresses are removed.
X-rayThe health view of the repo. See the X-ray.
SagasThe goals and their runs. See Sagas.
Skill libraryThe skills agents use, and the drafts waiting for your approval.
Code graph, System design, Design canvas, Design system, Authors, Constitution, Git graph, Issues, Notes, BrainstormThe other tools in the Studio, each opened from its tile.

Threads and worktrees

A thread is a line of work with its own git branch and its own worktree folder. Two threads can change the same file without clashing until you bring them together.

The Main thread is the workspace itself. Integrate, land and discard apply to the other threads.

Side questions

A side question is a read-only helper pane bound to one agent. Use it to ask about that agent's work, such as what it changed or why, without sending the agent a message. There is one side question per agent at a time.

3. Teams and graphs

A team is a graph of nodes and edges, edited on the board. Each run of the team is recorded in a saga, and each pass of a loop is its own episode in that saga.

The nodes

Explore the nodes

Pick a node to see what it does and what you set on it.

NodeWhat it doesWhat you set
LeaderYour terminal agent. It designs the team with you and watches the budget.Exactly one per graph.
AgentA planner, worker, QA, debugger, audit, or your own custom agent.Its brief, its skills, its tools, and the seat it runs on (engine:model).
ToolWeb search, page reading, or one of this app's own tools.An allow-list. The node may use only the tools it names.
ForkStarts branches that run side by side.The branches it starts.
JoinWaits for its branches.All of them, or a quorum.
If and switchRoute the run on what a node said.A condition. A switch needs a default port, so nothing falls through.
LoopRepeats its body until a condition holds.A max count and a cost cap. Each pass is its own episode.
For eachOne branch per item, in parallel up to a limit.The list it walks over, and the limit.
RetrySends a failed node round again.How many times, then the exhausted port.
Human gateWaits for your yes or no.A timeout, and the port taken when it runs out.
Cost capThe most a node, or a whole loop, may spend.A number in USD.
ArtifactA document the run makes, such as a report or a draft.A name and a format. It is written to .antiloki/artifacts/.
DeliverThe end of the graph. The finished pack reaches you.Nothing. Every graph ends in one.

What the validator refuses

Try the validator

Tick what your graph has, and see what the validator would refuse. This is a simplified model of its rules; the app checks more.

The validator checks the graph before anything runs. It refuses a graph that has any of these problems, and it names the problem so you can fix it on the board:

CodeWhat it means
no_leaderThe graph has no Leader.
many_leadersThe graph has more than one Leader.
no_nameThe graph or a node has no name.
no_entryNothing starts the graph.
no_deliveryThe graph has no Deliver, so the run has no end.
unreachableA node cannot be reached from the start.
dead_endA node leads nowhere and is not a Deliver.
edge_to_nowhereAn edge points at a node that does not exist.
bad_portAn edge uses a port the node does not have.
bad_conditionA condition on an if or switch cannot be read.
switch_no_defaultA switch has no default port.
loop_no_capA loop has no cost cap.
loop_unboundedA loop has no maximum count.
unbounded_cycleThe graph loops back on itself without a loop node that bounds it.
silent_failure_edgeA failure path goes nowhere, so a failure would disappear.

Artifacts and skills

An artifact is a document a run produces. Artifacts are written inside the project's .antiloki folder, and the next node in the graph reads them. You can open them from the run.

Agents can propose changes to their skills, the instructions and tools they carry. A proposal arrives as a draft in the skill library. Nothing changes until you approve it.

4. Budgets and gates

Money and writes are limited in three places:

Writes are fenced by folder. An agent can write only inside the folders its node is allowed to touch. A commit to your branch, a publish, or a message sent outside the project waits at a human gate for your yes or no.

Watch a run hit its budget

Costs are illustrative. The rule is the point: the run stops when the next step would pass the budget.

5. MCP

Antiloki is also an MCP server. Any MCP client can drive it: open threads and panes, read and add sagas, rebuild the X-ray, and design and run teams. Your agents use the same tools, so what an agent does in a pane is visible to you in the app.

The tools

Find a tool

    The names below are the tools the server registers today. Ask the client to list them to see the full set.

    FamilyTools
    Threads and spacesantiloki_threads, antiloki_spaces, antiloki_worktrees
    Panesantiloki_panes, antiloki_pane_open, antiloki_pane_read, antiloki_pane_send, antiloki_pane_view, antiloki_pane_act, antiloki_pane_arrange, antiloki_pane_close
    Teamsantiloki_teams, antiloki_team_design, antiloki_team_status, antiloki_team_simulate
    Sagasantiloki_sagas, antiloki_saga_add, antiloki_episode_add, antiloki_episode_update
    X-rayantiloki_xray, antiloki_xray_rebuild, antiloki_xray_open
    Delegation and checksantiloki_delegate, antiloki_qa, antiloki_rules, antiloki_query, antiloki_look
    Frames, screen and themesantiloki_frame_create, antiloki_screen, antiloki_themes

    Connecting a client

    Antiloki's MCP server runs with the app. Connection steps for each MCP client will come with the first download, so check back then.

    6. Web apps and the web node

    Web apps open in their own pane, from a store, beside your agents. You can keep the apps you use, and open one with a click.

    A graph has a web node for research. It works with no model and no key: it searches the web or reads one page, and the nodes after it are told what it found.

    What the web node does

    • Search order: a Brave key, if you have set one. Otherwise the seat's own web search. Otherwise DuckDuckGo, which needs no key.
    • Read only: it makes GET requests for text pages. It has a size cap and a timeout.
    • Address guard: it never reads an address inside this machine or its network, so a graph cannot be used to reach a router, a metadata service or the app itself.
    • Your own search: ANTILOKI_WEB_SEARCH_URL can name another search, as a template with {q}.

    7. The X-ray

    The X-ray is a health view of the repo, built from the code itself. The panels are below.

    The X-ray panels: verdict row, health over time, hotspots and the codebase map
    The X-ray: verdict row, health over time, hotspots and the codebase map. Example data, rendered from the site's own design. Not a capture of a live project.

    The verdict row

    FieldWhat it counts
    healthThe overall health score, out of 100. Higher is better.
    riskHow likely the next change is to break something.
    filesThe source files judged. Tests are not counted here.
    coverageThe share of executed lines, when a test run has measured it.
    solo-ownedThe share of files written mostly by one person. A high share is a knowledge risk.
    vulnerable depsDependencies with a known vulnerability, counted by severity.

    The panels

    The lenses

    Switch the lens

    Example values. Pick a lens to recolour the same boxes.

    A lens recolours the same map so that you can read a different question from it:

    LensQuestion it answersRed means
    healthWhich files are sick?Dense branching, code that is untested under load, and churn.
    cascadeIf this file breaks, what else breaks?Many files depend on it, through their imports, directly or not.
    ownershipWho knows this code?Recent commits come from a single author. That is a knowledge risk.
    debtWhat is waiting to be done in the code?TODO, HACK and unused markers. Ten or more is full red.
    heatWhat changes most?How often the file changed in the last 500 commits.
    runtimeWhat do the tests actually run?Grey means no runtime signal. Otherwise the colour shows how much of the file your tests execute.

    Each file has a "why" line in the lens view. It names the raw driver behind its colour, so a red box is never just a number.

    Rebuilding

    The X-ray reads the code as it was when it was last built. Rebuild it after large changes, or from the MCP tool antiloki_xray_rebuild. A rebuild re-reads the active workspace's own codebase.

    The chat in Cowork can read the X-ray, so you can ask about a score in plain words.

    8. Sagas

    A saga is a goal and the episodes that get it done. Sagas are listed per project, as a list or a board. Each row shows the progress, the cost, and when it last changed.

    The Sagas list, grouped by status
    Sagas, grouped by status. Example data, rendered from the site's own design. Not a capture of a live project.
    StatusMeaning
    todoNot started.
    in progressWork is running on it.
    doneFinished, with its record kept.
    failedA step in it failed.
    canceledStopped on purpose. Kept on the record.

    On the board you move a saga between columns. Moving a saga to done marks its open episodes done; moving it back to todo reopens it. A finished saga stays on the record and cannot be deleted, so the history is the history.

    A saga also keeps who did each step: the agent, the exact model, the time, and the cost, taken from the real run.

    9. Your AI seats

    A seat is one AI tool you are signed in to. Antiloki reads the seat, not the installed program, so a tool that is installed but not signed in is never offered.

    SeatNotes
    Claude CodeThe Leader's chat in Cowork is a Claude Code seat.
    Codex
    GeminiAntigravity is preferred when it is present, and Gemini is the fallback.
    Grok
    CursorThe agent that runs in the terminal, not the Cursor editor.
    OpenCodeOptional, off by default.

    One seat is enough. A team can use more than one, for example a planner on one and workers on another. A node names its seat as engine:model. A node that names no seat uses the one its tier picks.

    10. Install and status

    Downloads are coming soon for Windows, Linux and macOS. Nothing here needs an account.

    11. Privacy and network

    There is no account and no telemetry. The network calls are:

    12. Support

    Write to support@antiloki.com. Include the project's screen and the node or saga it is about, if you can.