Agent Tooling

ArcFlow runs in-process with zero setup. This makes it the natural operational world model for coding agents — spin up a spatial-temporal graph, process data with full graph algorithms and confidence scoring, query results, discard. No infrastructure to provision, no cleanup.

Why agents need an embedded graph#

When a coding agent is working on a task, it often needs to:

Explore relationships in data (who connects to whom, what depends on what)
Run algorithms (PageRank for importance, community detection for clustering)
Test queries before writing them into application code
Process structured data temporarily (parse a CSV, build a graph, extract insights)
Prototype a world model before committing to a schema

With a graph database server, the agent would need to ask the user to provision and start it. With ArcFlow, the agent just does it:

const db = openInMemory()  // instant, no permission needed
// ... process data ...
db.close()  // gone, no cleanup

Patterns#

Ad-hoc data exploration#

An agent analyzing a codebase can model dependencies as a graph:

import { openInMemory } from 'arcflow'
 
const db = openInMemory()
 
// Model file dependencies
db.batchMutate([
  "CREATE (f:File {path: 'src/index.ts', lines: 150})",
  "CREATE (f:File {path: 'src/utils.ts', lines: 80})",
  "CREATE (f:File {path: 'src/db.ts', lines: 200})",
  "CREATE (f:File {path: 'src/index.ts'})-[:IMPORTS]->(g:File {path: 'src/utils.ts'})",
  "CREATE (f:File {path: 'src/index.ts'})-[:IMPORTS]->(g:File {path: 'src/db.ts'})",
  "CREATE (f:File {path: 'src/db.ts'})-[:IMPORTS]->(g:File {path: 'src/utils.ts'})",
])
 
// Which files are most depended on?
const pr = db.query("CALL algo.pageRank()")
 
// Are there circular dependencies?
const components = db.query("CALL algo.connectedComponents()")
 
db.close()

Temporary world model#

An agent building a research summary can create a world model, query it, then discard it:

const db = openInMemory()
 
// Ingest extracted entities
db.batchMutate(entities.map(e =>
  `MERGE (n:${e.type} {id: '${e.id}', name: '${e.name}'})`
))
 
// Ingest relationships
db.batchMutate(relations.map(r =>
  `MATCH (a {id: '${r.from}'}) MATCH (b {id: '${r.to}'}) MERGE (a)-[:${r.type}]->(b)`
))
 
// Find the most connected entities
const important = db.query("CALL algo.pageRank()")
 
// Find clusters
const communities = db.query("CALL algo.louvain()")
 
// Get the answer, discard the graph
db.close()

Test fixture builder#

An agent writing tests can create graph fixtures inline:

import { openInMemory } from 'arcflow'
 
describe('social features', () => {
  it('finds mutual friends', () => {
    const db = openInMemory()
 
    db.batchMutate([
      "CREATE (a:User {name: 'Alice'})",
      "CREATE (b:User {name: 'Bob'})",
      "CREATE (c:User {name: 'Carol'})",
      "CREATE (a:User {name: 'Alice'})-[:FRIENDS]->(c:User {name: 'Carol'})",
      "CREATE (b:User {name: 'Bob'})-[:FRIENDS]->(c:User {name: 'Carol'})",
    ])
 
    const mutual = db.query(`
      MATCH (a:User {name: 'Alice'})-[:FRIENDS]->(m)<-[:FRIENDS]-(b:User {name: 'Bob'})
      RETURN m.name
    `)
    expect(mutual.rows[0].get('name')).toBe('Carol')
 
    db.close()
  })
})

Pipeline scratch space#

An agent processing data through multiple stages can use ArcFlow as a scratch graph:

const db = openInMemory()
 
// Stage 1: Load raw data
db.batchMutate(rawRecords.map(r =>
  `CREATE (n:Raw {id: '${r.id}', value: '${r.value}', source: '${r.source}'})`
))
 
// Stage 2: Deduplicate (find similar nodes)
const similar = db.query("CALL algo.nodeSimilarity()")
 
// Stage 3: Merge duplicates
// ... process similar pairs ...
 
// Stage 4: Extract result
const clean = db.query("MATCH (n:Raw) RETURN n.id, n.value")
const output = clean.rows.map(r => r.toObject())
 
db.close()
return output

Browser WASM — agent verification playground#

The browser runtime at oz.com/engine gives agents another tool: verify a query works by running it in the playground before writing it into code.

An agent can:

Open oz.com/engine (or instruct the user to)
Paste test data + query
Verify results
Then write the confirmed-working query into the application

With sync enabled, the playground becomes persistent:

oz.com/engine?sync=af_xxxxxxxxxxxx

The agent builds a graph in the browser, the user reviews it visually, then the same graph syncs to the application backend.

CLI binary — shell-native agents#

Shell-native agents call arcflow directly — the world model is a CLI tool, same invocation pattern as any other shell command. No protocol, no handshake, no session. The difference: every call returns from a confidence-scored, temporally-versioned graph, not a flat file:

# Symbol lookup — exits in <10ms
arcflow symbol login
 
# Impact traversal — what breaks if login() changes?
arcflow impact fn_login --depth 4
 
# Slice source at line range
arcflow slice src/auth.ts 12 35
 
# Run a GQL query directly
arcflow query 'MATCH (n:Function) RETURN n.name, n.line_start ORDER BY n.name'
 
# Git blame via the code graph
arcflow git-blame src/auth.ts

The agent calls these commands, reads stdout, and acts on structured results. No daemon to manage, no session state to track.

MCP — cloud chat interfaces only#

MCP is the right integration when the agent runs in a browser or cloud sandbox with no local shell — Claude.ai and similar cloud chat UIs. These interfaces cannot run arcflow as a binary, so MCP is the only way to give them graph tool access:

// Cloud chat UI configuration (Claude.ai, etc.) — NOT for Claude Code
{
  "mcpServers": {
    "arcflow": {
      "command": "npx",
      "args": ["arcflow-mcp"]
    }
  }
}

Available tools: get_schema, read_query, write_query, graph_rag. Latency is fine here — the user is already waiting for a chat response.

Why this matters for agent adoption#

Capability	Effect
`openInMemory()`	Agent can use a graph DB without asking user to start a server
Zero cleanup	Graph disappears on `close()` — no Docker containers to stop
Browser WASM	Agent can verify queries in a live playground
CLI binary	CLI agents query the graph directly via shell — `arcflow query`, `arcflow impact`
MCP server	Cloud chat UIs that have no local shell (Claude.ai, browser-only agents)
Sync	Ephemeral graph can optionally persist to cloud
Typed results	Agent doesn't have to parse strings — numbers are numbers
Structured errors	Agent can pattern-match on error codes and self-correct

ArcFlow is a tool agents can use directly — not infrastructure they have to ask the user to provision.

Agent Tooling

Why agents need an embedded graph#

When a coding agent is working on a task, it often needs to:

Explore relationships in data (who connects to whom, what depends on what)
Run algorithms (PageRank for importance, community detection for clustering)
Test queries before writing them into application code
Process structured data temporarily (parse a CSV, build a graph, extract insights)
Prototype a world model before committing to a schema

With a graph database server, the agent would need to ask the user to provision and start it. With ArcFlow, the agent just does it:

const db = openInMemory()  // instant, no permission needed
// ... process data ...
db.close()  // gone, no cleanup

Patterns#

Ad-hoc data exploration#

An agent analyzing a codebase can model dependencies as a graph:

import { openInMemory } from 'arcflow'
 
const db = openInMemory()
 
// Model file dependencies
db.batchMutate([
  "CREATE (f:File {path: 'src/index.ts', lines: 150})",
  "CREATE (f:File {path: 'src/utils.ts', lines: 80})",
  "CREATE (f:File {path: 'src/db.ts', lines: 200})",
  "CREATE (f:File {path: 'src/index.ts'})-[:IMPORTS]->(g:File {path: 'src/utils.ts'})",
  "CREATE (f:File {path: 'src/index.ts'})-[:IMPORTS]->(g:File {path: 'src/db.ts'})",
  "CREATE (f:File {path: 'src/db.ts'})-[:IMPORTS]->(g:File {path: 'src/utils.ts'})",
])
 
// Which files are most depended on?
const pr = db.query("CALL algo.pageRank()")
 
// Are there circular dependencies?
const components = db.query("CALL algo.connectedComponents()")
 
db.close()

Temporary world model#

An agent building a research summary can create a world model, query it, then discard it:

const db = openInMemory()
 
// Ingest extracted entities
db.batchMutate(entities.map(e =>
  `MERGE (n:${e.type} {id: '${e.id}', name: '${e.name}'})`
))
 
// Ingest relationships
db.batchMutate(relations.map(r =>
  `MATCH (a {id: '${r.from}'}) MATCH (b {id: '${r.to}'}) MERGE (a)-[:${r.type}]->(b)`
))
 
// Find the most connected entities
const important = db.query("CALL algo.pageRank()")
 
// Find clusters
const communities = db.query("CALL algo.louvain()")
 
// Get the answer, discard the graph
db.close()

Test fixture builder#

An agent writing tests can create graph fixtures inline:

import { openInMemory } from 'arcflow'
 
describe('social features', () => {
  it('finds mutual friends', () => {
    const db = openInMemory()
 
    db.batchMutate([
      "CREATE (a:User {name: 'Alice'})",
      "CREATE (b:User {name: 'Bob'})",
      "CREATE (c:User {name: 'Carol'})",
      "CREATE (a:User {name: 'Alice'})-[:FRIENDS]->(c:User {name: 'Carol'})",
      "CREATE (b:User {name: 'Bob'})-[:FRIENDS]->(c:User {name: 'Carol'})",
    ])
 
    const mutual = db.query(`
      MATCH (a:User {name: 'Alice'})-[:FRIENDS]->(m)<-[:FRIENDS]-(b:User {name: 'Bob'})
      RETURN m.name
    `)
    expect(mutual.rows[0].get('name')).toBe('Carol')
 
    db.close()
  })
})

Pipeline scratch space#

An agent processing data through multiple stages can use ArcFlow as a scratch graph:

const db = openInMemory()
 
// Stage 1: Load raw data
db.batchMutate(rawRecords.map(r =>
  `CREATE (n:Raw {id: '${r.id}', value: '${r.value}', source: '${r.source}'})`
))
 
// Stage 2: Deduplicate (find similar nodes)
const similar = db.query("CALL algo.nodeSimilarity()")
 
// Stage 3: Merge duplicates
// ... process similar pairs ...
 
// Stage 4: Extract result
const clean = db.query("MATCH (n:Raw) RETURN n.id, n.value")
const output = clean.rows.map(r => r.toObject())
 
db.close()
return output

Browser WASM — agent verification playground#

The browser runtime at oz.com/engine gives agents another tool: verify a query works by running it in the playground before writing it into code.

An agent can:

Open oz.com/engine (or instruct the user to)
Paste test data + query
Verify results
Then write the confirmed-working query into the application

With sync enabled, the playground becomes persistent:

oz.com/engine?sync=af_xxxxxxxxxxxx

The agent builds a graph in the browser, the user reviews it visually, then the same graph syncs to the application backend.

CLI binary — shell-native agents#

# Symbol lookup — exits in <10ms
arcflow symbol login
 
# Impact traversal — what breaks if login() changes?
arcflow impact fn_login --depth 4
 
# Slice source at line range
arcflow slice src/auth.ts 12 35
 
# Run a GQL query directly
arcflow query 'MATCH (n:Function) RETURN n.name, n.line_start ORDER BY n.name'
 
# Git blame via the code graph
arcflow git-blame src/auth.ts

The agent calls these commands, reads stdout, and acts on structured results. No daemon to manage, no session state to track.

MCP — cloud chat interfaces only#

// Cloud chat UI configuration (Claude.ai, etc.) — NOT for Claude Code
{
  "mcpServers": {
    "arcflow": {
      "command": "npx",
      "args": ["arcflow-mcp"]
    }
  }
}

Available tools: get_schema, read_query, write_query, graph_rag. Latency is fine here — the user is already waiting for a chat response.

Why this matters for agent adoption#

Capability	Effect
`openInMemory()`	Agent can use a graph DB without asking user to start a server
Zero cleanup	Graph disappears on `close()` — no Docker containers to stop
Browser WASM	Agent can verify queries in a live playground
CLI binary	CLI agents query the graph directly via shell — `arcflow query`, `arcflow impact`
MCP server	Cloud chat UIs that have no local shell (Claude.ai, browser-only agents)
Sync	Ephemeral graph can optionally persist to cloud
Typed results	Agent doesn't have to parse strings — numbers are numbers
Structured errors	Agent can pattern-match on error codes and self-correct

ArcFlow is a tool agents can use directly — not infrastructure they have to ask the user to provision.

Agent Tooling

Why agents need an embedded graph#

Patterns#

Ad-hoc data exploration#

Temporary world model#

Test fixture builder#

Pipeline scratch space#

Browser WASM — agent verification playground#

CLI binary — shell-native agents#

MCP — cloud chat interfaces only#

Why this matters for agent adoption#

See Also#

Agent Tooling

Why agents need an embedded graph#

Patterns#

Ad-hoc data exploration#

Temporary world model#

Test fixture builder#

Pipeline scratch space#

Browser WASM — agent verification playground#

CLI binary — shell-native agents#

MCP — cloud chat interfaces only#

Why this matters for agent adoption#

See Also#