The lean runtime for durable knowledge work

Agents that finish the work.

Build agents for work that crosses systems, files, people and time. Delta combines tools, memory, delegation, review and durable execution in one lean, inspectable runtime.

Explore the walkthrough Start building
$./dist/delta init ./my-agent
Durable hours of work across restart-safe checkpoints
Guarded recorded model usage checked between turns
Adaptive optional scoped memory and a versioned self-file
Inspectable thread, runs, files, data and setup in five views

Delta Cockpit

Inspect the work, not just the answer.

Open any thread to inspect every model turn, tool call, delegated task, file and recorded cost. These illustrative timelines use the same trace primitives and controls as delta dev.

Workflows
Renewal · churn · pricing · executive
Evidence
Model input · tools · journals · files
Explore
4 threads · 6 runs
Visibility
Tokens · cache · latency · cost
Delta Cockpit
agent account-intel tokens 92,157cost $0.0456runs 2 ILLUSTRATIVE DATA
tokens 92,157cost $0.0456runs 2
Acme renewal decision pack
Build an approval-ready Acme renewal decision pack from the CRM, 90 days of support, the account-team channel and inbox/2026-07-13/acme-call.txt. Challenge the narrative independently, save the brief and evidence note, then submit one proposal for review. Do not change the CRM.
Recall · 2 learnings(procedure/agent) Lead renewal briefs with material risk, then evidence.(preference/user) Preserve approved context when revising an existing review item.
Turn 1
input 6,840 tokens in · 82% cached
mcpcrm__get_account✓ 214ms
Journal · read-time redaction appliedargs { "account_id": "acme" } result { "arr_usd": 248000, "renewal": "2026-09-30", "probability": 0.62 }
mcpsupport__search_tickets✓ 486ms
Journal · read-time redaction appliedargs { "account": "acme", "window": "90d" } result 14 tickets · 4 open · permissions and data sync dominate
mcpslack__search✓ 362ms
Journal · read-time redaction appliedargs { "query": "acme renewal implementation" } result 18 days behind · sponsor engaged · no confirmed recovery owner
mcpreview__get_review_item✓ 188ms
Journal · read-time redaction appliedargs { "review_item_id": "RI-4821" } result { "revision": 2, "status": "changes_requested" }
built-inread_file✓ 24ms
Workspace readinbox/2026-07-13/acme-call.txt · 412 lines
anthropic/claude-sonnet-5 · out 161 · $0.0032 · 1,284ms · primary
Turn 2
input 8,112 tokens in · 86% cached
built-inwrite_file✓ 21ms
Workspace writewrote 3,842 chars to evidence/acme-source-pack.md
anthropic/claude-sonnet-5 · out 183 · $0.0035 · 1,176ms · primary
Turn 3
input 8,708 tokens in · 88% cached
built-inspawn_subagent✓ 6.8s
Fresh child context · shared workspace · narrowly forwarded model credentialtask Read evidence/acme-source-pack.md. Challenge the evidence independently. Identify contradictions, missing owners and unsupported claims. Save the note to evidence/acme-risk-check.md. Make no external changes. result saved evidence/acme-risk-check.md · 3 material risks · 2 contradictions · 1 missing owner
anthropic/claude-sonnet-5 · out 144 · $0.0037 · 1,288ms · primary
Turn 4
input 9,220 tokens in · 90% cached
built-inread_file✓ 18ms
Workspace readevidence/acme-risk-check.md · 2.7 KB
anthropic/claude-sonnet-5 · out 117 · $0.0033 · 1,046ms · primary
Turn 5
input 10,014 tokens in · 91% cached
built-inwrite_file✓ 22ms
Workspace writewrote 1,486 chars to briefs/acme-renewal.md
anthropic/claude-sonnet-5 · out 185 · $0.0042 · 1,338ms · primary
Turn 6
input 10,786 tokens in · 92% cached
mcpreview__propose_change✓ 274ms
Journal · proposal sent to connected review productargs { "supersedes_id": "RI-4821", "artifact_path": "briefs/acme-renewal.md", "run_ref": "resp_a3f09d2c8e174b65a2f941d6bc730e5f", "summary": "Revision 3 leads with material risk and includes an independent challenge." } result { "review_item_id": "RI-4821", "revision": 3, "status": "pending" }
anthropic/claude-sonnet-5 · out 92 · $0.0034 · 1,074ms · primary
Turn 7
input 11,700 tokens in · 93% cached
Built Acme's renewal decision pack. briefs/acme-renewal.md leads with the 18-day delay and unconfirmed recovery owner. evidence/acme-risk-check.md preserves the independent challenge. Revision 3 is pending on RI-4821. The CRM remains unchanged.
anthropic/claude-sonnet-5 · out 218 · $0.0078 · 1,422ms · primary
● done · run complete74,880 tokens · in 73,380 · out 1,500 · 90% cached · $0.0371
Review outcome for RI-4821 revision 3: accepted with one edit. Proposed: “no recovery owner.” Accepted: “recovery owner proposed; start date unconfirmed.” Reviewer note: distinguish proposed from committed ownership.
▶ started · review disposition
Turn 1
input 4,812 tokens in · 84% cached
mcpreview__get_review_item✓ 176ms
Journal · read-time redaction appliedresult { "review_item_id": "RI-4821", "revision": 3, "status": "approved_with_edit" }
built-inread_file✓ 16ms
Workspace readbriefs/acme-renewal.md · revision 3 proposed bytes
anthropic/claude-sonnet-5 · out 164 · $0.0026 · 924ms · primary
Turn 2
input 5,738 tokens in · 87% cached
built-inwrite_file✓ 19ms
Workspace writewrote 1,522 chars to briefs/acme-renewal.md
anthropic/claude-sonnet-5 · out 192 · $0.0028 · 988ms · primary
Turn 3
input 6,182 tokens in · 89% cached
Synced the accepted brief. It now distinguishes the proposed recovery owner from a committed owner and leaves the start date unconfirmed. Review item RI-4821 is approved. The CRM remains unchanged.
anthropic/claude-sonnet-5 · out 189 · $0.0031 · 906ms · primary
● done · run complete17,277 tokens · in 16,732 · out 545 · 87% cached · $0.0085
reflection · preference stored
Continue this thread...
composer preview · thread context preserved

Why Delta

Built for knowledge work beyond the repository.

Delta sits between raw model APIs and managed agent platforms: a self-hosted runtime for durable, cross-tool work.

Operating-model comparison Typical centers of gravity, not capability limits. Products vary. Scroll horizontally to compare
Typical centers of gravity for a model call, a coding harness, a hosted agent platform and Delta
Compare by PrimitiveModel call SpecialistCoding harness PlatformHosted agent platform RuntimeDelta
Starting point Model request Code repository Managed agent service Knowledge-work outcome
Center of gravity Generate one response Inspect and change code Build on managed primitives Complete cross-tool work
Runtime lives in Your application Your development environment The hosted platform Your infrastructure
Durability Added by your application Varies by harness Provided by the platform Built into the runtime
Best fit Bounded LLM features Repository-centered agents Managed agent applications Durable knowledge work
Delta boundary

Delta checkpoints runtime state. Your product owns identity, authorization, approval decisions, infrastructure lifecycle and consequential external side effects.

Use cases

One runtime. A chat assistant, or an agentic feature.

Both are the same agent — threads for context, tools for reach, review-driven reflection for learning. What changes is who opens the thread and where the human gives feedback.

01 · Agentic assistant

Chat that does the work, then gets better at it.

Opened by
A person, one thread per task
Work
Tools · permissioned app access · propose
Learns from
How each proposal is reviewed
02 · Agentic feature

Turn a one-shot LLM call into an agent that improves.

Opened by
Your product, seeded with task context
Work
Draft · propose · reflect on the diff
Learns from
Proposed versus accepted, per use case
03 · One runtime underneath

Same engine, threads, tools and review loop.

Config
One bundle per agent
Runtime
One loop · one memory model
Result
Two products, one thing to operate

The runtime

Files, MCP tools, memory and subagents. Built in.

Each capability has an explicit place in the prompt, runtime or durable state.

01

Work across tools and files

Research through MCP, work with files, delegate focused tasks and invoke a coding agent when needed.

FilesMCPSubagentsCode CLISchedules
02

Learn within bounds

Carry scoped memory, recall and bounded self-file updates into the next run.

DELTA.mdScoped memoryRecallReflection
03

Keep control explicit

Apply fixed policy, usage guards, checkpoints, cancellation and reviewed-write guidance.

PolicyUsage guardsReview railRecoveryRedaction

Bounded learning

Turn trusted feedback into scoped memory.

When enabled, successful runs can trigger best-effort reflection into scoped memory or a proposed shared capability. During a run, the explicit remember tool can update the bounded DELTA.md self-file.

Example learned rule + Lead renewal briefs with material risk, then evidence. scope agent · source review · confidence 0.96
  1. 01WorkRun the task
  2. 02ObserveKeep the trace
  3. 03ReviewCapture feedback
  4. 04DistillScope the lesson
  5. 05ReuseApply it next run

Control

Usage guards. Durable checkpoints. Human review.

Delta enforces runtime guards. Your product owns identity, approval and lifecycle.

01Usage guards

Recorded model steps, fresh tokens and cost are checked between turns.

02Durable checkpoints

Messages and tool journal entries persist. Interrupted side effects still require verification.

03Reviewed writes

Delta recognizes the rail; your product owns approval.

04Product boundary

Your gateway owns identity, tenancy, limits and lifecycle.

Long-running work

Runs for hours without dropping the thread.

A tool-heavy task can span dozens of steps and hundreds of thousands of tokens. Delta keeps the active window bounded and every earlier finding recoverable; nothing load-bearing is ever thrown away.

01Compact before overflow

The engine estimates each request and compacts older turns before sending, so a long run never blows the context window.

02Nothing is deleted

Compacted turns are archived, not dropped, and large results spill to disk with a pointer the agent can read back.

03Audited summaries

Each summary is checked to preserve every number, date and path, and merges forward across generations so a step-3 fact survives to step 50.

04Recall and recite

The agent searches its own compacted history with recall and keeps a working plan that rides through every compaction.

Knowledge-work patterns

From research to reviewable action.

Three patterns for work that crosses sources, artifacts, decisions and review.

01 · Account intelligence

Turn scattered signals into a review-ready decision.

Inputs
CRM · support · team notes
Work
Cross-check · synthesize · cite
Output
Renewal brief · review proposal
02 · Deep research

Parallelize the investigation, preserve one evidence trail.

Inputs
Web · files · MCP sources
Work
Subagents · evaluation · synthesis
Output
Sourced report · working files
03 · Controlled operations

Prepare the change, keep the consequential action reviewed.

Inputs
Schedule · events · product state
Work
Inspect · draft · checkpoint
Output
Reviewable proposal · durable trace

Build

Create and run in two commands. Shape four files.

Build once from source, then scaffold and launch a versionable agent bundle.

From source, use ./dist/delta. On your path, use delta.

Build once, then initialize and run

Install and compile from source. Then delta init scaffolds without overwriting files, and delta dev opens the local Cockpit.

Terminal 
# build the binary
bun install
bun run build

# create and launch an agent
./dist/delta init ./my-agent
./dist/delta dev ./my-agent

Shape four files. Add context if needed.

The bundle is deliberately plain. Version it, review it and change it without learning a framework.

delta.envProviders, budgets and MCP
vocab.jsonProduct language and actions
DELTA.mdRole, memory and learned rules
POLICY.mdHighest-priority prompt guidance
PROMPT_CONTEXT.md optionalBounded stable and per-turn context

Inspect the real loop locally

delta dev runs the ordinary daemon on loopback, adds exact successful-call capture and local editing, and opens Cockpit at /dev.

Runtime telemetry

Trace runs, tools, tokens and cost.

Correlate each event with its user, agent, session, run and turn. Inspect live, persist locally or export NDJSON.

Read the telemetry setup
Event stream anatomy custom NDJSON · GenAI-inspired fields
Correlation fieldsIncluded when present
  1. 01user.id
  2. 02agent.id
  3. 03session.id
  4. 04run.id
  5. 05task.id
  6. 06entity.id
  7. 07turn
Example eventmodel.call
A single envelope keeps the operational context together.

See the served model and tracked usage without opening the full prompt.

Model
anthropic/claude-sonnet-5
Tokens
4,812 total
Cost
$0.0148
Latency
1,184 ms
  1. 01
    Inspect liveCockpit and task SSE
  2. 02
    Persist locallyRestart-stable SQLite records
  3. 03
    Export as NDJSONBatched, retried while retained and deduplicable
TELEMETRY_URL=https://collector.example/ingestTELEMETRY_TOKEN=your_bearer_token
Export contract DELTA_CAPTURE_PAYLOADS=1 preserves per-call model and tool attributes. It never exports prompt or result bodies; exact call capture stays local to Cockpit.

Deploy

One binary. One volume. Your cloud.

Pair one Delta binary with one persistent volume. Your controller handles intake, wake and suspend. Delta checkpoints the work.

Read the deployment guide
Production topology one agent · externally enforced writer · persistent state
External Lifecycle controller Gate intake · check queue · wake · suspend
Private Trusted gateway and operator proxy Authenticate · apply limits · inject control and inspect tokens
Delta machine suspendable compute
delta:8080

Compiled daemon, durable loop and Cockpit

  • /v1/responses
  • /healthz
  • /dev
  • /v1/dev/*
/data/delta.db SQLite state and durable event outbox
/data/workspace/ DELTA.md · inbox/ · working files
  • ModelModel providersOpenRouter, Anthropic or OpenAI
  • ToolsMCP servicesProduct systems and connectors
  • OptionalTelemetry collectorCustom NDJSON export
  • BackupLitestream + object storageSQLite only · snapshot the workspace separately
External lifecycle Delta does not provision or suspend infrastructure itself.
  1. 01GateHold new work
  2. 02WakeStart the runtime
  3. 03RunCheckpoint durable work
  4. 04CheckNo queued or running work
  5. 05SuspendCompute sleeps, /data stays

Queue zero does not drain reflection or telemetry. Suspending can interrupt background work.

Runtime
Compiled Linux binary
Persistence
Mount all of /data
Health
/healthz is liveness
Recovery
Back up DB + workspace; test coordinated restore

Provider neutral

Switch providers. Keep the same agent contract.

Use compatible Chat Completions, native Anthropic Messages or OpenAI Responses while preserving policy, identity and tools.

01

OpenRouter

Route across models with streaming, provider choice and cost-aware usage.

DEFAULT
02

Anthropic

Use native Messages with prompt caching and configurable thinking budgets.

NATIVE
03

OpenAI

Use native Responses, or compatible Chat Completions through the OpenAI-compatible route.

RESPONSES
04

Codex subscription

Broker Responses access for approved internal deployments; tokens are restricted to configured allowlisted hosts.

INTERNAL PREVIEW

Coding agent integrations

Codex and Claude Code native.

At launch, Delta will hand advanced coding tasks to either CLI with the agent workspace as its working directory, then return the result to the run.

  1. Delta
  2. code(task)
  3. CLI
  4. Resume

  • Codex

    delta.env DELTA_CODE_CLI=codex exec --sandbox workspace-write --skip-git-repo-check

  • Claude Code

    delta.env DELTA_CODE_CLI=claude --print

Install either CLI separately. Neither ships in the standard Delta image.

  • Working directoryAgent workspace
  • AuthenticationCLI-owned credentials
  • Environment variablesNarrow allowlist
  • ExecutionNo harness timeout

Canonical guide

Build, inspect, deploy and recover from one guide.

The canonical operating guide is written for engineers and language models.

Start an agent Read the docs