Teams · Agents · Mesh

Your enterprise is an ecosystem.Orchard brings your AI into it.

Orchard connects an ecosystem of AI agents with your people, teams and tools. Set direction in Orchard Chat; agents take on the work and bring results back. Coordinate swarms across sessions and machines, reuse their output as shared input, and spend less on repeated work.

Explore the mesh Talk about your team

Claude Code + Codex Connect with MCP

Services & BPO

Put Orchard to work
for your business.

Modernize what you run. Build what comes next. Bring us a defined project or an ongoing workstream, and we’ll shape the scope, coordinate delivery and keep your team in the loop.

Engineering projects Ongoing delivery & BPO
Migrate

Lift and shift legacy applications

Move existing systems to a new environment with a clear cutover plan.

Dependency mappingCutover planRollback
Modernize

Refactor legacy codebases

Make inherited software easier to maintain, test, and extend.

Codebase assessmentRefactoringRegression checks
Create

Develop new applications

Take a new product or internal system from brief to working software.

Product definitionDevelopmentRelease
Keep moving

Ongoing delivery & BPO

Give recurring software and infrastructure work a team, a clear scope, and a review cadence.

Scoped backlogDelivery reviewsShared knowledge

The economics of swarming

Stop paying twice
for the same work.

Models bill by the word, going in and coming back. Sprawling context, rambling answers, duplicated investigation, overwritten changes and builds that started before the approach was agreed all send the same words again. Orchard’s job is to make that repetition rare. Put a number on it with your own workload.

One source, up closeA change gets erased. The work is bought again.
  1. 01 / CREATE

    Agent A writes

    A change lands in a 4,000-token file.

  2. 02 / COLLIDE

    Agent B overwrites

    An overlapping edit removes A’s work.

  3. 03 / DISCOVER

    Tests fail

    The agent rereads code, diffs and logs.

  4. 04 / REPEAT

    Pay to repair

    Diagnose the failure. Generate the fix again.

164,000billed tokens in this worked example

150,000 in + 14,000 out, across the several steps it takes to find the problem and redo the work: code and test rereads, diagnostics, reasoning and a regenerated file. Both agents’ intended edits are excluded. This is one of six sources priced in the calculator below.

Build the business case

What will Orchard save you?

See what removing repeated work could save your team. Model spend only — no staff time, no productivity claims.

Illustrative model · your assumptions · USD

Your numbers

Estimated monthly savings$3,811.71
$45,740over twelve identical months

About 20% less modeled AI spend for 25 people. After 5% coordination overhead. Orchard pricing is not included.

Where the savings come from

Removing 50% of each selected source of repeated work. Monthly savings before coordination costs.

  • Sending more than the job needs$1,856.25
  • Answers longer than the question$515.63
  • Re-reading work to check it$1,237.50
  • The same problem solved twice$268.13
  • One change erasing another$136.13
  • Building before the approach is agreed$519.75
Repeated spend avoided
$4,533.38
Less coordination overhead
-$721.67
Net monthly savings
$3,811.71

Category amounts are rounded; totals use unrounded values.

See the full bill comparison
Your AI bill, without Orchard
$18,967
With 50% of the repetition removed
$15,155
  • Work you wanted
  • The same work, bought again
  • Coordination overhead
Modeled monthly AI spend, USD
CategoryYour AI bill, without OrchardWith 50% of the repetition removed
Work you wanted$9,900.00$9,900.00
The same work, bought again$9,066.75$4,533.38
Coordination overhead$0.00$721.67

Both bars use the same dollar scale. The work you wanted does not move: doing the job once costs the same either way. Only the repeated half is in play.

Modeled AI spend / person / month$759Repeated share of that bill48%Billed tokens avoided / month734.7M

Check the modeled bill against a real invoice before you trust the rest. If it is too high or too low, change the tasks per day until it matches.

Scope a pilot against your own baseline

Six ways the same job gets bought twice

Providers charge by the word, both for what you send and for what comes back. These are the places the same words get sent again. Amounts here show repeated spending before any savings. Untick anything you do not recognise in your own team to update the savings above.

  • $3,713repeated / month

    Every step drags along sprawling files and old conversation the task never uses. You are billed for all of it, on every step.

    Smaller files · shared memorySmall, single-purpose files and a shared record of what was already learned, so a step carries the part that matters.

    Every task · 8,250 times a month · 742.5M billed tokens

  • $1,031repeated / month

    Restated background, narration and padding. Generated text is the most expensive thing you buy, and this is the part nobody asked for.

    Tight responsesResponse rules that return the decision and the change instead of describing the work along the way.

    Every task · 8,250 times a month · 41.3M billed tokens

  • $2,475repeated / month

    With no automated check, the model is asked again to confirm its own work, re-reading the same material to answer a question a test could settle.

    Test plans · CI gatesA test plan travels with the change and the pipeline runs it on every push, so the answer comes from CI instead of another paid round.

    Every task · 8,250 times a month · 396M billed tokens

  • $536repeated / month

    Two workers investigate the same thing in parallel. Both read the same code, both reason it through, and you pay for both.

    CoordinationAssigned scopes and a shared record of findings, so the second one starts from the first one’s result instead of repeating it.

    5% of tasks · 413 times a month · 87.5M billed tokens

  • $272repeated / month

    A second change lands on top of the first and removes it. It surfaces later, and the whole investigation and fix is bought again.

    No overwritingDeclared ownership and a check before writing, so two workers never edit the same file blind.

    3% of tasks · 248 times a month · 40.6M billed tokens

  • $1,040repeated / month

    A full build runs on the wrong approach and is thrown away. The bill covers the discarded attempt and the real one.

    Proposal first · predict then implementA short written proposal and a predicted result, checked while changing course still costs a paragraph instead of a build.

    7% of tasks · 578 times a month · 161.7M billed tokens

Priced at published list rates for the model you selected. None of this is a measured Orchard result: it is your own workload, costed against the practices Orchard enforces.

Inspect the math, pricing and assumptions

Monthly opportunity = (people × tasks per day × working days) × the share of tasks each source touches × its repeated tokens at list price, × the share you expect to remove, − what coordination costs you.

Tokens assumed repeated, per affected task. Every figure is editable.
Source of repeated work% of tasksRepeated input tokensRepeated output tokens
Sending more than the job needs
Answers longer than the question
Re-reading work to check it
The same problem solved twice
One change erasing another
Building before the approach is agreed
Standard API rates · USD per million tokens · checked 2026-09-17
ModelInputOutputCache read
Claude Sonnet 5$2.00$10.00$0.20
Claude Opus 5$5.00$25.00$0.50
GPT-6 Astra$10.00$50.00$1.00

The default task is a multi-step agent session, not a single question: an agent reads, runs something, reads the result and continues, and every step re-sends the context so far. That is why per-task token counts run to the hundreds of thousands. Tokens are priced at uncached list rates, so the estimate is conservative for teams already getting cache hits; provider caching is a separate saving. GPT-6 Astra uses its short-context tier. Equal token counts compare prices, not model quality or equal work.

These are API-equivalent costs. On a subscription, avoiding work may preserve your usage allowance without changing the invoice. Annual figures repeat the same month twelve times; they are not a forecast. Staff time, tools, infrastructure, taxes and provider discounts are outside the model, and nothing here is a measured Orchard result.

Download this scenarioRead the methodology
One migration. Three workstreams.
Generate useful output once

A reviewed migration plan

A capable model maps the constraints. The team checks the plan and shares it.

Reuse the findings as input
Implementation

Shared plan as input

Update the code

Start with the findings. Spend effort on the assigned task.

Validation

Shared plan as input

Check the behavior

Start with the findings. Spend effort on the assigned task.

Documentation

Shared plan as input

Update the runbook

Start with the findings. Spend effort on the assigned task.

Shared reasoning gives each agent a head start. Models can be chosen to fit each task.

Illustrative workflow. Savings depend on the work avoided, model pricing and the context carried forward.

Output becomes input

Generate once. Reuse the result.

Keep useful findings in shared memory, rules and skills. The next agent reads the relevant result as input instead of spending output tokens rediscovering and explaining the same thing.

Coordinated swarming

Divide the work. Avoid doing it twice.

Give agents distinct scopes, clear ownership and a shared starting point. Run independent tasks in parallel, share discoveries and bring the results together for review.

Model subsidization

Let stronger reasoning support cheaper work.

Invest in a capable model for a difficult decision. Reuse its reviewed plan as context for lower-cost or local models on scoped tasks. One model’s work supports the rest of the swarm; your team chooses the models and checks the results.

The cost shift: less repeated generation, more useful context.

The same 4,000 tokens.
A different cost to use them.Illustrative unit comparison · Claude Opus 5
Generate again as output$0.10
Read the existing result as input$0.02

80% lower token cost for this reuse step. Other task costs still apply.

Input is typically priced below output, and eligible provider prompt caching can reduce the cost of repeated input further. Orchard makes useful outputs available to reuse; provider caching is a separate saving. Compare the full cost of an accepted result, including context, coordination and review.

Choose the model for each part of the work.

Configure Chat and worker models separately, with hosted and local options.

Hosted models

Choose the model behind Chat.

DeepSeek + configured providers

Use the hosted models enabled for Orchard Chat. Your team can choose its coordination model independently of the agents doing the work.

About Chat model selection

Chat has separate Text, Thinking and Files selections. The models available depend on your configuration and provider access.

Worker sessions

Set up the agents for the task.

Harness-specific model settings

Dispatch through supported runtimes such as Codex and Claude Code. Worker sessions use their own model settings; changing Chat’s model does not change theirs.

About worker model settings

The agent runtime and its configured dispatch defaults determine the worker model. Include those sessions, their tools and review work when assessing total cost.

Local inference

Run compatible models on your hardware.

Local runtime + model of your choice

Use compatible local models through a configured runtime. Hosted DeepSeek access and an open-weight model running locally are different deployment choices.

About local model requirements

Local Chat needs a ready runtime and a model that supports its required capabilities, including tool use. Hardware capacity, setup and operation still have costs.

Compare the whole job.

Include model usage, repeated context, parallel sessions and review in the cost of an accepted result.

Explore recorded usage
How we measure

Team coordination

See the work between agents.

Follow who exchanged work, which tools and memory they used, and how their sessions connect. Explore an actual workspace record, with each route open to inspection.

Recorded workspace graph13 Sep 2026 · 06:48 UTC
37recorded participants
10active at capture
354directed sends · 24h
73directed message routes
Services → agents → correspondents
100%
01 / Infrastructure
02 / Active at capture
03 / Recorded correspondents

Scroll or swipe inside the graph to pan. Select a node to inspect its routes.

Follow an event through OrchardDrill into the record, routing and receipts behind the mesh.

Recorded capture, anonymized. One observed machine. Messages: 12–13 Sep 2026, 06:48 UTC. On message routes, dot density follows the recorded message count. Memory and action routes carry one dot and claim no volume. Timing is not measured latency.

Capture scope and limits

352 local ledger rows (up to 8 × 500); 354 local mirrored talk rows (limit 500), deduplicated by message ID. Both message reads reached the end of their window. One ledger row had no resolvable sender and was excluded.

354 sends across 73 directed routes. Actions use the newest 500 rows in the same requested window; more rows exist. Identity enrichment used a 200-row project roster and the active local roster. Unknown runtimes remain unknown. This is a bounded observation, not a complete fleet inventory.

Memory shows positive recorded operation counts. Trellis and Cellar counts overlap. Dispatch origins show one recorded conversation-routing relationship, not a completed dispatch. Dashed machine routes show observed placement. Anonymized aliases are arbitrary; no task role is inferred.

The shared record

One shared record, whichever tool they run.

The graph above shows which sessions exchanged work. This is what moved between them. Every prompt, tool call, permission decision and handoff is written to an append-only record before it is delivered. Claude Code, Codex and Gemini land in the same shape, so you read one history instead of one per vendor.

Coordination eventsanonymized capture · one workstation
17:02:46agent-04Tool callpre_tool_use · Bash · git · allowed
17:02:46agent-04Promptuser_prompt_submit · session s_4KpR
17:00:30agent-11Sessionsession_end · session s_9TvM
17:00:22agent-11Tool callpost_tool_use · Bash · coord · allowed
17:00:08agent-11Messageagent_message/sent · to s_4KpR
16:59:15agent-02Tool callpost_tool_use · Read · read · allowed
16:59:14agent-11Tool callpost_tool_use · apply_patch · edit · allowed
16:59:07orchardSessionruntime_context/acked · session s_4KpR
16:58:52agent-04Tool callpost_tool_use · Bash · search · allowed
16:58:41agent-07Promptuser_prompt_submit · session s_2QdX
16:58:30agent-02Tool callpre_tool_use · Write · edit · allowed
16:58:18agent-11Messagebeam_coordination/delivered · from s_9TvM
16:58:05agent-04Sessionsession_start · session s_4KpR

The highlighted row, in full

One agent addressing another. The record keeps who acted, where, on what, and whether it was allowed.

operator
codexwhich runtime acted
session_id
s_9TvMthe session that acted
target_session_id
s_4KpRwho it was addressed to
repo
api-gateway @ mainwhich repository, at which commit
activity_intent
coordwhat it was trying to do
decision
allowedwhether it was permitted
event_hash
5b9d92ef48c7…content digest

Fifty-seven fields are recorded per event. Sample rows, shortened for reading; identities, hosts and paths are replaced with placeholders.

See the record audited, and a session correcting another

Case studies

See Orchard at work.

Explore real projects through the work, the reviews and the results.

All case studies
The Stratus airspace-intelligence application in the browser

Stratus · Built with an agent swarm

Four hours from idea to airspace demo.

One hour of human work. More than 80 combined agent-hours. The creator’s account of a four-hour build, with a look inside the application today.

Inside the Stratus build
MATHEMATICSFrom the research record
cα=1S∑0<β≤αwβcα−β.(2)\boxed{c_\alpha=\frac1S\sum_{0<\beta\le\alpha}w_\beta c_{\alpha-\beta}.} \tag{2}

S = total degree of α > 0 · c₀ = 1

HardyRecurrence.leanLean checked
theorem coefficients_nonneg
    (c w : Index → ℝ)
    (hc0 : c 0 = 1)
    (hw : ∀ β, 0 ≤ w β)
    (hrec : ∀ α : Index, α ≠ 0 →
      (degree α : ℝ) * c α =
        ∑ β ∈ (box α).erase 0, w β * c (α - β)) :
    ∀ α, 0 ≤ c α := by
Conditional result: the starting value, nonnegative weights and recurrence are assumptions.

Math research · Ongoing

AI agents doing real mathematics.

Read the equations, inspect the Lean proofs, and follow the research with a plain-English TL;DR.

Latest research updateAnother proof step passes Lean, with its assumptions explicit
Follow the research
Rendered agent and coordinator prototype models from the Blender redesign
Prototype models · Blender redesign

Design & development

Rebuilding the 3D coordination graphic.

From rejected concepts to custom Blender models and an interactive browser scene.

Explore the redesign
A cloud and server illustration created for the use-case website
Illustration from the website project

Software delivery

Building a website with an agent team.

Explore the working site and the build record: scoped tasks, review and verified corrections.

Explore the recorded build

Security & audit

Inspect the record, its attribution, and its gaps.

The record above is not only for you. Each coordination event is committed before it is delivered, with the attribution its runtime supplies, and kept where a later rule can read it. Coverage views help identify windows where the pipeline was not receiving.

  1. CommittedThe occurrence is written before the event is delivered.
  2. AttributedAccount, machine and session, where the runtime reports them.
  3. Findings integrityVerification checks finding fields, predecessor links and an anchor.
  4. ReplayableA rule written later can still read it.
CoverageIllustrative shape — your workspace reports its own

Recorded windows, and the one in the middle where the pipeline was not receiving. A gap is stored as a gap rather than closed over, because a record that hides its blind spots is the one an auditor cannot use.

Attribution reaches a person

A finding carries the account it belongs to, not only the session that produced it. Sessions are ephemeral; the person answering for the work is not.

It records where it was not looking

Absence gets a table of its own. The record reports the windows the audit pipeline was not receiving, so a reviewer can see where it is incomplete instead of assuming it is whole.

A new rule reads old events

With captured history on disk, a rule written today runs backwards over events captured long before it existed. You are not limited to what you thought to watch for at the time.

Evidence, not assurance

Which agent touched this repository, under whose account, on what machine. Orchard produces the record an auditor asks for. What it proves depends on the coverage it reports alongside it.

The operating model

Your organization is the team.
AI is part of it.

Connect the people who own the outcome with the agents, knowledge and machines doing the work.

People direct the work

One conversation to direct the work.

Your team sets the brief in Orchard Chat. Delegate work to agent sessions, follow up as they run and bring their reports back into the conversation. Open a session’s terminal when you need to inspect the work.

  • A central chat for dispatch and follow-up
  • Separate sessions for implementation and review
  • Session reports return to the conversation
Follow a reviewed build
Orchard ChatWalkthrough · sample content

Start with the team’s objective.

Reports return here

Select the brief or a session to look inside.

Team direction → agent work → team review

Teams · Knowledge · Mesh

One ecosystem for your teams and their agent tools.

  • Claude Code
  • Codex CLI
  • Grok Build CLI
  • OpenCode
  • Hermes Agent

Trellis · memory & token reuse

Remember the work.
Reuse the context.

Trellis gives agents a place to recall previous work. See the memory, token usage and reuse recorded in the workspace where we build Orchard.

Memory and model usage in Orchard

Product screenshot
Open image (new tab)
Supplied dashboard: estimated memory savings 44.6M tokens and 2.9 days of model compute; 1,008,441 operations; 348,054 records. All-time usage: 176,450 model calls, 918M consumed tokens, 25.9B cache-read prompt tokens and 155,422 calls across the top 15 tool names.
Our own workspace · June to September 2026. Memory and usage share this capture. The 30-day label applies to the cache graph, not the all-time usage tiles. This capture is cropped at its right edge.

Recall instead of re-deriving.

Keep findings available to the next agent. The saved history below estimates the repeated work avoided through recall.

The screenshot’s 44.6M-token estimate spans 3.3M–158M. Its 2.9 days estimate is model compute, not human time. These are modeled savings, not measured reductions.

One total. Three components.

918M tokens

All-time consumed tokens · rounded values from the snapshot

Fresh input
476M
Cache writes
346M
Output
96M
Cached prompt reuse, shown separately25.9B cache-read tokens

97% prompt cache share. This does not represent a 97% reduction in cost.

One workspace’s recorded usage, not a benchmark or a typical result. Provider caching and Trellis recall are separate mechanisms.

Read the methodology

Usage & reuse

See where repeated context can be reused.

Explore the saved history

Savings over time

Aug 14, 2026, 12:50 – Sep 13, 2026, 06:45 UTC · 2,799 saved observations

Estimated tokens avoided through recall44.3M tokens in range
Bars · 0–4MDashed cumulative · 0–44.3M
Estimated tokens avoided through recallBars show estimated tokens per 6 hour interval. Dashed line shows cumulative tokens within the selected range on its own scale. Inspect exact values with the slider below.
Aug 14, 2026, 12:50 to Aug 14, 2026, 18:50 UTC: 1,260,000 tokens; cumulative 1,260,000 tokens
View exact interval values
Estimated tokens avoided through recall · UTC · 6 hour groups
Interval starttokensCumulative
Aug 14, 2026, 12:501,260,0001,260,000
Aug 14, 2026, 18:50980,0002,240,000
Aug 15, 2026, 00:50130,0002,370,000
Aug 15, 2026, 06:5002,370,000
Aug 15, 2026, 12:50470,0002,840,000
Aug 15, 2026, 18:50230,0003,070,000
Aug 16, 2026, 00:50500,0003,570,000
Aug 16, 2026, 06:5040,0003,610,000
Aug 16, 2026, 12:50410,0004,020,000
Aug 16, 2026, 18:5040,0004,060,000
Aug 17, 2026, 00:50480,0004,540,000
Aug 17, 2026, 12:50900,0005,440,000
Aug 17, 2026, 18:50380,0005,820,000
Aug 18, 2026, 00:50640,0006,460,000
Aug 18, 2026, 06:5006,460,000
Aug 18, 2026, 12:50360,0006,820,000
Aug 18, 2026, 18:50230,0007,050,000
Aug 19, 2026, 00:50100,0007,150,000
Aug 19, 2026, 12:50280,0007,430,000
Aug 19, 2026, 18:501,530,0008,960,000
Aug 20, 2026, 00:501,840,00010,800,000
Aug 20, 2026, 06:50600,00011,400,000
Aug 20, 2026, 12:503,200,00014,600,000
Aug 20, 2026, 18:503,600,00018,200,000
Aug 21, 2026, 00:50200,00018,400,000
Aug 21, 2026, 12:50800,00019,200,000
Aug 21, 2026, 18:50900,00020,100,000
Aug 22, 2026, 00:50020,100,000
Aug 22, 2026, 12:50400,00020,500,000
Aug 22, 2026, 18:50400,00020,900,000
Aug 23, 2026, 00:50600,00021,500,000
Aug 23, 2026, 06:50100,00021,600,000
Aug 23, 2026, 12:50300,00021,900,000
Aug 23, 2026, 18:50400,00022,300,000
Aug 24, 2026, 00:50100,00022,400,000
Aug 24, 2026, 12:50700,00023,100,000
Aug 24, 2026, 18:504,000,00027,100,000
Aug 25, 2026, 00:50100,00027,200,000
Aug 25, 2026, 12:503,000,00030,200,000
Aug 25, 2026, 18:50700,00030,900,000
Aug 26, 2026, 00:501,100,00032,000,000
Aug 26, 2026, 12:502,100,00034,100,000
Aug 26, 2026, 18:50100,00034,200,000
Aug 27, 2026, 00:50034,200,000
Aug 27, 2026, 12:50100,00034,300,000
Aug 27, 2026, 18:50034,300,000
Aug 28, 2026, 00:50034,300,000
Aug 28, 2026, 12:50200,00034,500,000
Aug 28, 2026, 18:50100,00034,600,000
Aug 29, 2026, 00:50034,600,000
Aug 29, 2026, 18:50034,600,000
Aug 30, 2026, 12:50100,00034,700,000
Aug 30, 2026, 18:50034,700,000
Aug 31, 2026, 00:50100,00034,800,000
Aug 31, 2026, 12:50900,00035,700,000
Aug 31, 2026, 18:501,000,00036,700,000
Sep 1, 2026, 00:50200,00036,900,000
Sep 1, 2026, 12:50100,00037,000,000
Sep 1, 2026, 18:50037,000,000
Sep 2, 2026, 12:50300,00037,300,000
Sep 2, 2026, 18:50100,00037,400,000
Sep 4, 2026, 12:50037,400,000
Sep 4, 2026, 18:50037,400,000
Sep 5, 2026, 00:50037,400,000
Sep 5, 2026, 06:50037,400,000
Sep 5, 2026, 12:50100,00037,500,000
Sep 5, 2026, 18:50037,500,000
Sep 6, 2026, 00:50037,500,000
Sep 6, 2026, 06:50037,500,000
Sep 6, 2026, 12:50200,00037,700,000
Sep 6, 2026, 18:50100,00037,800,000
Sep 7, 2026, 00:50037,800,000
Sep 7, 2026, 12:50100,00037,900,000
Sep 7, 2026, 18:50100,00038,000,000
Sep 8, 2026, 00:50038,000,000
Sep 8, 2026, 12:50038,000,000
Sep 8, 2026, 18:50038,000,000
Sep 9, 2026, 00:50038,000,000
Sep 9, 2026, 06:50038,000,000
Sep 9, 2026, 12:50038,000,000
Sep 9, 2026, 18:50200,00038,200,000
Sep 10, 2026, 00:50038,200,000
Sep 10, 2026, 12:50400,00038,600,000
Sep 10, 2026, 18:50900,00039,500,000
Sep 11, 2026, 00:50500,00040,000,000
Sep 11, 2026, 06:50500,00040,500,000
Sep 11, 2026, 12:50040,500,000
Sep 11, 2026, 18:50700,00041,200,000
Sep 12, 2026, 12:501,300,00042,500,000
Sep 12, 2026, 18:50900,00043,400,000
Sep 13, 2026, 00:50900,00044,300,000
Observed record increases339.6K records in range
Bars · 0–24KDashed cumulative · 0–339.6K
Observed record increasesBars show recorded records per 6 hour interval. Dashed line shows cumulative records within the selected range on its own scale. Inspect exact values with the slider below.
Aug 14, 2026, 12:50 to Aug 14, 2026, 18:50 UTC: 7,465 records; cumulative 7,465 records
View exact interval values
Observed record increases · UTC · 6 hour groups
Interval startrecordsCumulative
Aug 14, 2026, 12:507,4657,465
Aug 14, 2026, 18:506,99314,458
Aug 15, 2026, 00:501,73416,192
Aug 15, 2026, 06:50116,193
Aug 15, 2026, 12:503,14319,336
Aug 15, 2026, 18:501,75521,091
Aug 16, 2026, 00:503,96625,057
Aug 16, 2026, 06:5044125,498
Aug 16, 2026, 12:501,66427,162
Aug 16, 2026, 18:5063027,792
Aug 17, 2026, 00:504,98432,776
Aug 17, 2026, 12:504,36237,138
Aug 17, 2026, 18:501,87039,008
Aug 18, 2026, 00:505,74544,753
Aug 18, 2026, 06:50044,753
Aug 18, 2026, 12:501,40046,153
Aug 18, 2026, 18:501,35647,509
Aug 19, 2026, 00:5045247,961
Aug 19, 2026, 12:501,57249,533
Aug 19, 2026, 18:5014,45663,989
Aug 20, 2026, 00:5010,93374,922
Aug 20, 2026, 06:504,58779,509
Aug 20, 2026, 12:5016,45795,966
Aug 20, 2026, 18:5020,639116,605
Aug 21, 2026, 00:503,203119,808
Aug 21, 2026, 12:5023,957143,765
Aug 21, 2026, 18:503,896147,661
Aug 22, 2026, 00:50317147,978
Aug 22, 2026, 12:501,591149,569
Aug 22, 2026, 18:502,309151,878
Aug 23, 2026, 00:506,135158,013
Aug 23, 2026, 06:50301158,314
Aug 23, 2026, 12:501,598159,912
Aug 23, 2026, 18:501,492161,404
Aug 24, 2026, 00:50548161,952
Aug 24, 2026, 12:502,723164,675
Aug 24, 2026, 18:5020,726185,401
Aug 25, 2026, 00:501,195186,596
Aug 25, 2026, 12:5021,719208,315
Aug 25, 2026, 18:5012,188220,503
Aug 26, 2026, 00:5020,078240,581
Aug 26, 2026, 12:5012,592253,173
Aug 26, 2026, 18:502,543255,716
Aug 27, 2026, 00:50468256,184
Aug 27, 2026, 12:501,924258,108
Aug 27, 2026, 18:50686258,794
Aug 28, 2026, 00:50327259,121
Aug 28, 2026, 12:50433259,554
Aug 28, 2026, 18:50699260,253
Aug 29, 2026, 00:5016260,269
Aug 29, 2026, 18:50138260,407
Aug 30, 2026, 12:50313260,720
Aug 30, 2026, 18:50225260,945
Aug 31, 2026, 00:50539261,484
Aug 31, 2026, 12:506,278267,762
Aug 31, 2026, 18:5015,199282,961
Sep 1, 2026, 00:507,591290,552
Sep 1, 2026, 12:50892291,444
Sep 1, 2026, 18:504291,448
Sep 2, 2026, 12:502,446293,894
Sep 2, 2026, 18:50807294,701
Sep 4, 2026, 12:5093294,794
Sep 4, 2026, 18:501,833296,627
Sep 5, 2026, 00:50331296,958
Sep 5, 2026, 06:500296,958
Sep 5, 2026, 12:50771297,729
Sep 5, 2026, 18:50508298,237
Sep 6, 2026, 00:501,786300,023
Sep 6, 2026, 06:50565300,588
Sep 6, 2026, 12:501,313301,901
Sep 6, 2026, 18:501,111303,012
Sep 7, 2026, 00:50496303,508
Sep 7, 2026, 12:502,208305,716
Sep 7, 2026, 18:502,540308,256
Sep 8, 2026, 00:50563308,819
Sep 8, 2026, 12:501,988310,807
Sep 8, 2026, 18:501,671312,478
Sep 9, 2026, 00:50117312,595
Sep 9, 2026, 06:50546313,141
Sep 9, 2026, 12:50173313,314
Sep 9, 2026, 18:502,167315,481
Sep 10, 2026, 00:50440315,921
Sep 10, 2026, 12:503,225319,146
Sep 10, 2026, 18:503,337322,483
Sep 11, 2026, 00:502,198324,681
Sep 11, 2026, 06:501,370326,051
Sep 11, 2026, 12:5012326,063
Sep 11, 2026, 18:50797326,860
Sep 12, 2026, 12:507,967334,827
Sep 12, 2026, 18:503,769338,596
Sep 13, 2026, 00:50984339,580

Bars show counter increases recorded in each 6-hour group; observation gaps are not interpolated. Dashed lines accumulate those increases within this range, using their own scales. Record increases are not completed tasks or the current store population. No counter drops in this range.

Saved browser observations captured Sep 13, 2026, 06:50 UTC. Recall savings are modeled avoided work, not measured human time or provider cache usage.

Saved workspace snapshots · not liveDownload savings observationsDownload cache aggregates
Snapshot provenance

Savings: saved Orchard browser counter observations, retained for up to 30 days. Avoided tokens are estimates; observed record-counter increases are separate. Provider cache: Orchard Engine’s recorded usage aggregates in 15-minute buckets over the preceding 30 days. Cache timestamps refer to ledger insertion and boundary buckets can be partial. Both snapshots contain only aggregate counts and timestamps. Their windows and capture times differ, and their totals are never added together. Coverage depends on recorded telemetry.

The operator workspace

Give each contributor the tools to act.

Connect the operating model to the work on each machine. Contributors bring supported agent runtimes, a browser and configured project tools into their workspace.

Browser + agentProduct screenshot
Orchard desktop with its browser next to a Codex terminal.
The browser and a coding agent, together in an Orchard project. Open full-size screenshot
Multiple runtimesProduct screenshot
OpenCode, Grok Build, and Codex displayed side by side inside Orchard.
Separate agent runtimes in the same workspace. Shown settings belong to this captured session. Open full-size screenshot

Plugins and integrations

Orchard connects your agents to the tools your team already uses.

Messaging platforms your agents can speak through, and the apps you sign in to and pin inside the workspace. Discord, WhatsApp, Slack, Outlook, Teams, Gmail, and the rest of Google Workspace and Microsoft 365.

  • Discord
  • WhatsApp
  • Slack
  • Telegram
  • Signal
  • Microsoft Teams
  • iMessage
  • Matrix
  • LINE
  • Outlook
  • Gmail
  • Google Calendar
  • Google Drive
  • Google Docs
  • Google Sheets
  • Google Slides
  • Google Meet
  • Word
  • Excel
  • PowerPoint
  • OneNote
  • OneDrive
  • Jira
  • Stripe
  • Supabase
  • ChatGPT
  • Claude

Seedpacks bundle MCP servers, so a session can reach tools beyond this list.

FAQ

Questions people ask

Go deeper into the architecture
Is Orchard just another coding agent?

Orchard connects your team’s agent sessions, tools, memory and machines. Orchard Chat is a central place to delegate work, follow up with sessions and receive their reports, while people retain responsibility for the outcome.

Why use Orchard if I already use Claude Code or Codex?

Because work spans sessions, tools and teammates, and an agent loop does not. For one small task in one session, your agent is already enough.

Does using more agents make work cheaper or faster?

Not automatically. Coordinated swarming can reduce duplicated work: agents take distinct scopes, share findings and reuse earlier output as input. That can avoid paying several models to repeat the same investigation or plan. The benefit depends on task fit and the cost of coordination, context and review; adding agents alone does not guarantee savings.

What do you mean by model subsidization?

One model’s work supports the others. A capable model can produce a reviewed plan or resolve a hard question; lower-cost or local models can use that result as context for suitable follow-up tasks. Orchard helps carry the knowledge between them. Your team still selects the models and verifies the result.

Do all runtimes work the same way?

Tool access, hooks and permissions differ between runtimes. Orchard Chat’s model choices are separate from worker runtime and model settings. Hosted options such as DeepSeek also differ from compatible models running through a local runtime.

What happens to context, access, and permissions?

Rules, skills and memory are plain files on your machine that you can read and delete. Where data goes depends on the provider and integrations you configure.

Start with your team

Bring your teams into one AI ecosystem.

Start with a shared project. Connect your people, agents and tools, then extend the work across your connected machines.
Free during the beta.

Open Orchard Connect to Claude Read the setup guide

Remote MCP is in early access. Keep Orchard running and signed in on your computer. MCP setup and supported apps

Planning a team rollout?Talk with the Orchard team
Or start from your terminal
npm install -g @orchard-ai/cli

Use the setup guide to connect your workspace.