Aller au contenu principal

Do Everything From Chat

Every dashboard page in Ever Works carries a chat rail, and the rail is not a help widget. It calls the same REST API the buttons call, signed in as you, and it can create a Mission, queue an Idea for build, assign a Task to an Agent, enable a plugin, chart your spend, or generate a comparison — one entity at a time, asking before anything irreversible.

This guide is the hands-on companion to Platform Chat & Canvas. That page describes the system; this one gives you ten prompts that work, says exactly which tool each one calls, and tells you where in the dashboard to check the result.

Routes are written without the locale prefix — the address bar shows /en/works, this guide says /works.

Step 1 — Open the rail and pick a provider​

  1. Sign in and open any dashboard page — /works is a good one because most examples below want a Work in reach.
  2. Click the robot icon on the right edge of the navigation sidebar. The panel slides in. Its open/closed state is stored in the chat-panel-open cookie, so it is already open the next time you land on a dashboard page.
  3. In the toolbar, open the Provider selector on the right. It lists your AI-provider plugins. One showing a Not configured badge is disabled until you add credentials at /settings/plugins → AI Providers; without a configured provider the composer says "This provider is not configured. Set it up in Plugins." and a send attempt fails with "Unable to send message".
  4. Optionally pin a model with the Model picker beside the composer. It offers Provider default first, then the provider's configured tier models (simple / medium / complex), then a searchable list of the provider's full catalogue, fetched only when you open the picker. The pin is stored on the conversation row, so it survives a reload.
  5. Type into the composer (placeholder Ask me anything…). Enter sends, Shift+Enter adds a line, Stop generating aborts a streaming reply.
Say the domain out loud

The rail has roughly 400 tools, and no provider will accept 400 function schemas per turn. Each turn activates an always-on core plus every domain whose keywords appear in your last three messages or in the current page URL, capped at 128 tools. So say the noun: agent, task, mission, idea, plugin, webhook, budget, comparison, knowledge. If the assistant claims it cannot do something, name the domain in your next message and its tools are pulled into that turn. Nothing is lost permanently.

Two more things the rail does on its own, and it helps to expect them:

  • It uses where you are. The current page URL travels with every message, so on /works/:id/… you never paste the Work id.
  • It moves the page. After creating a Work it navigates you to the new Work; after other mutations it reloads the page so the dashboard shows what just changed.

Step 2 — Ten prompts that do real work​

Each row is a prompt you can paste, the tool it lands on, and where to verify it. Every tool named here is registered in the live tool set — the generated single-entity registry (apps/web/src/lib/ai/tools/generated/registry*.ts) or the hand-written tools beside it.

#Say thisTool that runsCheck it at
1Create a mission that researches AI developer tools every Monday at 09:00createMission/missions
2Add an idea: a directory of open-source observability toolscreateIdea/ideas
3Build that ideabuildIdea/ideas/:id
4How much has the Researcher agent spent this period?list_agents → get_agent_budget/agents/:id/budgets
5Show my blocked tasks as a tablelist_tasks → renderTable/tasks
6Assign that task to the Researcher agent and start itassign_task_to_agent/tasks/:id
7Enable the Slack connectorlist_plugins → enable_plugin/plugins
8Attach the Observability Directory work to that mission as createdattachWorkToMission/missions/:id
9Generate the next comparison for this workgenerate_comparison/works/:id/generator/comparisons
10List this work's knowledge-base documents and show me the brand voice onelist_kb_documents → showComponent/works/:id/kb

1. Create a Mission​

Create a mission that researches AI developer tools and drafts a weekly digest. Run it every Monday at 09:00 and keep at most five unbuilt ideas.

createMission takes a description (at least 10 characters — it becomes the Mission's AI Goal context), an optional title, a type of one-shot or scheduled, a five-field cron schedule, autoBuildWorks, and outstandingIdeasCap. The tool defaults to one-shot and only flips to scheduled when you explicitly ask for recurring runs, so "every Monday at 09:00" is what turns your sentence into type: "scheduled", schedule: "0 9 * * 1". autoBuildWorks is never set unless you ask for it — the Mission spawns Ideas and waits.

If you leave out the cadence, the assistant asks rather than inventing one. That is a rule it works under, not a quirk: missing name, cadence or URL is always a question.

Verify: /missions lists the new Mission; /missions/:id has its Run now, Pause, Clone and budget controls. See Missions.

2. Add an Idea​

Add an idea: a directory of open-source observability tools, with pricing and self-host notes for each entry.

createIdea needs only a description; the server derives the title when you omit one. Attachments added through a composer's + button travel with the prompt as upload ids and are attached to the Idea.

Verify: /ideas. See Ideas.

3. Build it​

Build that idea.

buildIdea transitions the Idea from PENDING (or FAILED) to QUEUED, spawns a Work build request under the hood, and returns the build-request id so the assistant can point you at the live run. Ask instead for refreshIdeas ("find me new ideas") to run the research job that proposes fresh ones — it is rate-limited server-side and reports status: "rate-limited" rather than failing.

If you already have a Work that fulfils the Idea and only want the back-reference, say so: that is acceptIdea, which links Idea to Work without generating anything.

Verify: /ideas/:id, then the Work it produces under /works.

4. Spend, per Agent​

How much has the Researcher agent spent this period?

Two calls: list_agents to resolve the name to an agent, then get_agent_budget (GET /api/agents/{id}/budget) for that agent's current-period spend. Ask for it as a dial — "show that as a gauge" — and the assistant follows with showComponent using the gauge component, which is built for budget and cap usage.

Spend questions scale up and down from there:

AskWhat runs
"What has this work spent, by plugin?"runReport → work_spend_by_plugin
"Chart this work's daily spend"runReport → work_spend_trend
"What have I spent across everything?"runReport → account_spend_overview
"What did task T-12 cost?"get_task_spend
"Cap this work at $50 a month"create_work_budget (needs cap + currency)

Verify: /agents/:id/budgets for the Agent, /works/:id/settings/budgets-usage for the Work, /settings/usage for the account. See Budgets & Usage.

5. List blocked Tasks​

Show my blocked tasks as a table.

list_tasks accepts status and missionId query filters, so this is one call with status: "blocked", followed by renderTable to put the rows in the canvas rather than in the message. Ask for "my tasks as a board" instead and you get runReport → tasks_board, a kanban artifact grouped by status.

Verify: /tasks, which has the same status, priority, scope, label and search filters. See Tasks.

6. Run a Task with an Agent​

Assign the "Refresh the pricing table" task to the Researcher agent and start it.

assign_task_to_agent (POST /api/agents/{id}/assign-task, body taskId) is the tool that actually dispatches work. It returns a runId, and it de-duplicates: if that Agent already has an in-flight run for that Task, you get the existing run back instead of a second one. The run goes through the same per-Work concurrency valve as every other dispatch, so a busy Work may return the run as queued with a reason rather than starting it immediately.

Related tools in the same breath: transition_task moves a Task's status (in_progress, completed, blocked), add_task_assignee / add_task_reviewer / add_task_approver set people on it, and post_task_chat writes into the Task's own thread.

Verify: /tasks/:id — the run appears in the Task's activity and run history. See Tasks and Agents.

7. Enable a plugin​

Enable the Slack connector.

list_plugins finds the plugin id, enable_plugin turns it on for the account, and get_plugin_connection_status tells you whether it is actually reachable — enabling a connector is not the same as authenticating it. Scope it to one Work instead with enable_work_plugin.

Turning a plugin off is confirmation-gated (disable_plugin, disable_work_plugin); turning one on is not.

Verify: /plugins for the catalogue, /plugins/:pluginId for its settings, /settings/plugins for the per-category view, /works/:id/plugins for the Work-scoped list. See Plugins and Integrations.

8. Attach a Work to a Mission​

Attach the Observability Directory work to the AI developer tools mission, as the work it created.

attachWorkToMission records a typed relation — one of created, improves, operates, markets, researches, retires — between a Mission and an existing Work. It never transfers or changes ownership: a Mission does not own Works, and the same Work can relate to many Missions across its life, even to the same Mission under several relation kinds. The assistant resolves both ids first with listWorks and listMissions, then calls attachWorkToMission; listMissionWorks reads the relations back and detachWorkFromMission removes exactly one edge.

Verify: the Attached Works panel on /missions/:id. See Missions.

9. Generate a comparison​

Generate the next comparison for this work.

Open /works/:id/generator/comparisons first so the Work id comes from the URL. generate_comparison auto-picks the next best pair of items and generates the page; get_comparison_generation_status reports progress on an in-flight run and get_remaining_comparison_count says how many pairs are left. For a specific pair, say which two — that routes to generate_manual_comparison, whose body takes itemASlug and itemBSlug.

Deleting one (delete_comparison) is confirmation-gated.

Verify: /works/:id/generator/comparisons. See Comparisons.

10. Ask the Knowledge Base​

List this work's knowledge-base documents, then show me the brand voice one.

On a Work page the assistant already knows which Work you mean. list_kb_documents reads that Work's Memory, and rendering a document is showComponent with the markdown component so the body lands in the canvas rather than flooding the transcript. create_kb_document (path, title, content) and update_kb_document write back; delete_kb_document is confirmation-gated. list_kb_tags and create_kb_tag manage the tag set.

When a reply references a document with a kb:{class}/{slug} token — kb:brand/voice, say — a Cited: footer appears under the message with one hover chip per document, resolved through the same endpoints the Memory workbench uses.

Verify: /works/:id/kb. See Knowledge Base & Memory.

Confirmation cards, and how to cancel​

Anything that deletes, removes, revokes, disconnects, cancels, rotates a secret, or spends on your behalf is marked as needing confirmation. 54 of the 332 generated tools carry the flag — delete_task, revoke_api_key, remove_work_member, rotate_webhook_secret, disconnect_oauth, cancel_generation, run_agent_now, send_email_message, delete_comparison, leave_work and their siblings.

The gate lives in the tool factory, not in the model's judgement. Called without confirmed: true, a gated tool returns a marker instead of touching the API — no request is made — and the chat renders a card:

  1. The card is titled Confirm this action and reads "<what the tool does> (<target>). This can't be undone."
  2. Click Confirm. The card collapses to a Confirming… spinner and sends a message naming the exact tool and target, so the model re-issues that call with confirmed: true. The spinner is deliberately a pending state, not a success claim — the mutation runs after the model processes the message.
  3. Click Cancel instead and the card collapses to Cancelled, having sent "No, cancel <tool> — do not proceed." Nothing was called.
  4. Changed your mind after Confirm? There is no undo — that is the point of the card. Recover the way you would from the dashboard: recreate the entity, or restore from Git for anything that lives in your repository.

The card names the tool and its target for a reason: if two confirmations are pending at once, confirming one cannot be misread as confirming the other.

The no-bulk guard​

There is no bulk anything. Bulk endpoints are excluded from the registry, and the factory rejects any call whose arguments or body carry an array with more than one entry — regardless of the field name. Ask for "delete all my works" and you get:

Bulk operations are not allowed in chat. Multiple values were supplied for "ids". Please ask me to do this one entity at a time.

Then the assistant asks which single entity to start with. A single-element array is fine.

Reading the canvas, reports and History​

The canvas​

When a tool produces something better seen than read, it is rendered into the Canvas, a slide-over panel on the right, and the chat shows a compact chip ending in · in canvas. Click the chip to focus that artifact. When a conversation has produced several, a tab strip appears above the panel; Close canvas dismisses it without losing anything, because artifacts are saved with the conversation.

ArtifactToolWhat you see
chartrenderChartLine, bar, area or pie, one or more series
tablerenderTableA scannable grid — works, items, agents, tasks, runs
statrenderStatCardsA row of metric tiles with optional hints
detailrenderDetailOne entity's fields plus status badges
kanbanrunReportColumns of cards, e.g. tasks grouped by status
componentshowComponentA named component — gauge, progress, timeline, markdown, funnel, kpi

Canvas tools never ask for confirmation: they only draw data that was already fetched.

Reports​

Analytics questions have a turnkey path. Ask the question in English and runReport fetches, aggregates and renders in one call.

AskReport id
"How are my tasks distributed?"tasks_by_status
"Show my tasks as a board"tasks_board
"How many agents are active?"agents_by_status
"Chart my missions by status"missions_by_status
"Spend over time for this work"work_spend_trend
"This work's spend by plugin"work_spend_by_plugin
"Activity per day"activity_per_day
"Items generated per day for this work"work_items_per_day
"Members of this work by role"work_members_by_role

Say "what reports are there?" to get listReports. When nothing named fits, buildReport groups any of nineteen list sources — tasks, agents, missions, ideas, works, skills, notifications, webhooks and deliveries, plugins, organizations, notification channels, API keys, templates, and the Work-scoped items, members, KB documents, comparisons and deployments — by any field into a bar or pie chart. "Chart this work's items by category" needs no named report.

Work-scoped reports need a Work. Run them from a /works/:id/… page and the id comes from the URL; run them from elsewhere and the assistant will ask which Work you mean.

History​

History in the toolbar lists your saved conversations, newest first, dated Today, Yesterday, 3d ago, then a plain date. Click a row to reopen the thread with its messages — and its canvas artifacts — intact; hover a row for the delete icon. New chat starts a fresh thread and carries your pinned model over.

Everything History shows is also available over REST at /api/conversations, which is worth knowing when you want to script an export. See AI Conversation.

From Slack​

With the Slack connector enabled, you get the same engine in a channel.

You doWhat happens
@works what shipped today?The mention is routed into the platform chat as the bound user and the answer is posted back into that thread.
/works what shipped today?You get an instant private acknowledgement — "On it — I am asking Ever Works now…" — and the answer is posted into the channel when it is ready.
/works with nothing afterA usage hint naming the command and an example question, rather than an empty prompt sent to the model.
Either, from a workspace nobody has connected"This Slack workspace is not connected to an Ever Works account yet. Enable the Slack connector in Ever Works and try again."

Both paths verify Slack's request signature, resolve which install owns the workspace from the delivery's team id, and use your configured AI provider and model.

Slack answers, it does not operate

The Slack bridge runs a plain completion — it does not carry the dashboard's tool loop, so it has no confirmation cards and no canvas. "Delete the staging webhook" typed in Slack is a question about a webhook, not a deletion. Do the operating from the rail.

See Integrations.

From any OpenAI-compatible client​

The engine behind the rail is exposed at POST /api/v1/chat/completions, so any OpenAI-style client can talk to your platform, with your provider plugins and your Work context.

Header / fieldWhat it does
Authorization: Bearer …Required. Acts as you.
x-work-idScopes the completion to one Work — per-Work AI plugin settings apply, and @kb:{class}/{slug} mentions in the latest user message resolve to that Work's Memory.
x-provider-overrideRoute this request to a specific AI-provider plugin id.
modelA model id, or auto for the provider's configured default.
stream: trueReturns text/event-stream with chat.completion.chunk frames and a final data: [DONE].
curl -X POST "https://api.ever.works/api/v1/chat/completions" \
  -H "Authorization: Bearer $EVER_WORKS_TOKEN" \
  -H "x-work-id: $WORK_ID" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "auto",
    "messages": [
      { "role": "user", "content": "Summarize @kb:brand/voice in three bullets." }
    ]
  }'

With no AI provider configured the endpoint answers 422 { "error": { "type": "provider_unavailable" } } — a truthful 4xx, never a 500. Tools you pass in the body are handed to the model and come back as tool_calls; executing them is the caller's job. The rail is one such caller, and its tool loop is what turns "delete the staging webhook" into a confirmation card rather than a suggestion.

Omit x-work-id and the platform falls back to your first Work, so pass it whenever the answer depends on which Work you mean.

Chat or MCP?​

The MCP server exposes a curated whitelist of API operations to external AI assistants over the Model Context Protocol, authenticated with an API key. It is the right surface when the conversation is happening outside Ever Works. It is a deliberately smaller one.

CapabilityDashboard chat railMCP server
Tool surface~400 toolsA curated whitelist of operations
Runs asYour logged-in sessionAn API key
Confirmation card before destructive actionsYesNo — the client's own approval UI is all there is
No-bulk guard enforced in the tool layerYesNot applicable — bulk routes are not whitelisted
Canvas artifacts (charts, tables, boards, gauges)YesNo
Built-in reports (runReport, buildReport)YesNo
Knows which page you are onYes — the URL travels with every messageNo
Navigates and reloads the dashboard after a changeYesNo
Conversation history with artifacts, replayableYesNo
Works inside your editor or an external assistantNoYes

Use MCP to reach Ever Works from wherever you already are. Use the rail when you want the confirmation cards, the canvas, and a dashboard that keeps up with what you just did.

When something does not work​

SymptomCause and fix
"This provider is not configured. Set it up in Plugins."No credentials on the selected AI-provider plugin. Add them at /settings/plugins → AI Providers, or pick a different provider in the toolbar.
The assistant says it cannot do something you know it canPer-turn tool gating dropped that domain. Name the noun — "the webhook for staging", "this agent" — and its tools return on the next turn.
It keeps asking for an id you can see on screenYou are not on the entity's page. Open /works/:id/… (or the Mission, Task, Agent page) and ask again; the URL is passed with every message.
A report answers "needs a workId"It is a Work-scoped report. Run it from a Work page, or name the Work in your message.
"Bulk operations are not allowed in chat."By design. Ask for one entity, then the next.
A destructive request produced a card and then nothingThe card was left unanswered, so the tool never ran. Click Confirm or ask again.
The page still shows stale data after a changeThe rail reloads after mutations, but a reload that raced a slow write can miss it. Refresh the page.
422 provider_unavailable from the APISame root cause as the first row, seen from POST /api/v1/chat/completions.