Skip to main content
BitsWeave
Onboarding

Explore context

Use the /app/context explorer to traverse your org's context graph from a focal set and stream a prose summary — then call the same context API manually for structured output.

Once an agent is connected, its work builds up a context graph in your org: activity rows, ranked idents, and the relations between entities. The context explorer at /app/context lets you pick a focal set and see the graph plus a live prose summary. The same data is available over the context API for agents and scripts.

What you're looking at

Context is assembled from your org's procedure log + identity graph. You start from a focal set — one or more idents (a project, an agent, a member, …) — and BitsWeave traverses outward to find what co-occurred and what's related. See the MCP surface for how agents reach the same data.

Using the explorer (/app/context)

Open the explorer in the app:

Open the context explorer

It's a form over a view panel. The whole view is URL-driven — the focal set, depth, and budget live in the query string, so any walk you build is shareable and bookmarkable. Pasting a link reproduces the same walk.

Build a focal set

Each row in the Focal set form is one ident: a Kind dropdown plus a value control.

  • For graph-backed kinds — Project, Agent, Member, User, Note, Decision — the value control is an autocomplete: start typing and pick a match.
  • For free-text kinds — Organization, Agent run, Inbox item — type the id directly.

Add more rows with Add ident to seed a multi-ident walk; remove rows you don't want. At least one completed ident is required before you can explore.

Organization isn't a focal kind

The form lists Organization, but af-api rejects org as a focal ident — your active org is already implicit from your membership. If you pick it, the explorer surfaces a clean error rather than running the walk. Use a concrete entity (project, agent, …) as your focal seed instead.

Tune depth and budget

Two knobs shape the walk:

  • Depth (0–3, default 1) — how many hops out from the focal set to traverse.
  • Budget (1–1000, default 200) — the cap on ranked + activity rows returned.

Explore

Press Explore. The form submits as a GET — the params write to the URL, the loader runs context.traverse against them (org resolved server-side from your session pin), and the view panel renders the resulting graph (nodes + edges) below. If the walk fails (e.g. a forbidden focal kind, or idents that span multiple orgs), you get a clean error banner instead of the graph.

Read the summary

Alongside the graph, the explorer streams a prose summary of the same focal set — a plain-language read of what's going on, delivered incrementally as the model produces it (server-sent deltas keyed off the same URL state).

Using the context API manually

The explorer is a thin client over af-api's context.* procedures. Agents call them as MCP tools; you can also hit the RPC endpoint directly. All of them are gated on the context:read scope, and the org is never a parameter — it's inferred from your org-bound idents (or your active-org binding for context.graph).

ProcedureShapeUse it for
context.traversequery (buffered)The full envelope for a focal set: activity rows + ranked idents + hydrated entities + a pre-projected graph. The explorer's graph.
context.traverseStreamstreamingSame walk, streamed: one activity frame per row, then a terminal done frame.
context.summarystreamingThe prose summary — content deltas as plain strings. Not an agent tool (it would recurse on itself in-loop).
context.hydratequeryJust the per-entity snapshots for a focal set — no walk, no graph.
context.graphqueryThe org-wide constellation: newest entities of every kind as nodes + relation edges. No focal set.

The traverse/summary input mirrors what the form posts:

  • focalIdents — array of { kind, id } (1–50). org is rejected as a kind.
  • depth — 0–3 (default 1).
  • budget — 1–1000 (default 200).

A direct call to the structured context.traverse envelope:

curl -X POST 'https://bitsweave.com/api/rpc/context.traverse' \
  -H 'Content-Type: application/json' \
  -H "Authorization: Bearer $AGENTFLOW_TOKEN" \
  -d '{
    "json": {
      "focalIdents": [{ "kind": "project", "id": "<project-id>" }],
      "depth": 1,
      "budget": 200
    }
  }'

The response envelope carries activity, ranked (each with score plus recencyScore / frequencyScore / relationScore sub-scores and a viaPath), entities (hydrated focal + ranked), a pre-projected graph (nodes + edges with viaRelation labels), and relationDepth.

From an agent, prefer traverse over summary

context.traverse is the structured-output surface agents should call in their tool loop — it returns data, not prose. context.summary is for a human-facing read (the explorer's summary pane), and is deliberately not exposed as an agent tool because an agent calling it from inside its own dispatch path would recurse on itself.

Walks can be bounded out

A dense graph at high depth can hit the traverse timeout — af-api returns a BAD_REQUEST telling you to lower depth or narrow the focal set rather than hanging or 500-ing. Start at depth 1 and widen only if you need to.

Where this fits

The explorer and the API are the read side of everything your connected agent produces. With an org joined, an agent connected, and context flowing, you've completed onboarding — explore the Concepts and Guides next.

Connect your agents

Connect the AI agents you already run to BitsWeave in three steps — add the MCP server over OAuth, install af-tap with the af-tap-install skill, then verify by creating an artifact and watching a live session.

What is BitsWeave

Connect the coding agent you already use to a shared platform that tracks your work, gives it a memory, and lets it act across your tools.

On this page

Using the explorer (/app/context)Build a focal setTune depth and budgetExploreRead the summaryUsing the context API manuallyWhere this fits