Self-hosted AI workforce

Stop building agents. Start delegating work.

Install SRW on your Kubernetes cluster, choose a portable expert, and hand it a job. You get back a reviewable result, a complete audit trail, and an itemized receipt.

Install in minutes Watch a week of work
verified_userRuns on your infrastructure receipt_longEvery job metered codeFair Source, free to self-host
Delegate
Job 8d604690 · Finance Created
Job
Expert
Connectors
Workspace
historyAudit trail
    Pending review
    Duration
    Tokens
    Cost
    Audit steps
    187tools in 27 categories
    12connector types
    16model families, 6 providers
    115MCP tools for your IDE
    12 + 15experts and skills bundled
    EN · DEcockpit languages
    Check it out yourself in the repository arrow_forward

    Delegate a project

    A week of work in ninety seconds.

    One project, one note to the officer, and a legion of experts doing the rest. Press play, or step through it.

    Project Market analysis · Kinzig Verpackung GmbH scheduleMon 09:00 I / VI

    I

    Give the order once.

    A project, a standing officer with a roster and a daily ceiling, and one note that carries the whole brief.

    Experts

    An expert is a file.

    Adapt it, version it, and share it with another team. Its persona, tools, model, and subagent roster live in one YAML file. No Python, no workflow builder.

    Pick one. Read it. Change it.

    A bespoke agent will always beat a shared expert on the one task you spent a thousand hours on. You cannot spend a thousand hours on every task. A good expert file, adapted in an afternoon, is ninety percent as good on ninety percent of them, and every improvement you make is one you keep.

    Every expert extends the same base and runs in every role: as a one-off job, as a session you talk to, or as a subagent another expert calls. Swap the model in one line.

    touch_appClick an expert to open its file.

    Experts8 experts · project Finance
    descriptionexperts/critic.yamlvalidv12
    
          
    
        
    
      
    
    
    
    
    
      
    
        
    Bring your own model
    
        
    SRW does not resell tokens. Point an expert at the model you already pay for, or at one you run yourself. Per expert, per phase, per subagent, with a separate cheap model for summaries.
    
      
    
      
    
        
    OpenAIllm.modelgpt-5.6-sol
    
        
    Anthropicllm.modelclaude-opus-5
    
        
    Googlellm.modelgemini-3-pro
    
        
    Mistralllm.modelmistral-large
    
        
    MiniMaxllm.modelminimax-m3
    
        
    OpenRouterprovideropenrouter
    
        
    Your ownbase_urlvllm.internal/v1
    
      
    
      
    check_circleNo keys resold, no quota capscheck_circleSwitch the model in one line of the expert filecheck_circleThe receipt shows which model did the work
    
    
    
  
    

    



    
  
    
    
    
      
    Connectors
    
      
    Attach any data source. Or all of them.
    
      
    Repositories, databases, cloud folders, mailboxes, knowledge bases, MCP servers, SSH hosts. Attach one to a project and every expert in it can use it. Read-only or read-write, private or shared.
    
    
    
    
    
      
    
        
    Attached at run time.
    
        
    Connectors are tested when you add them and re-indexed when they change. Add one now and the next job sees it. Credentials stay with the orchestrator and are issued to a worker only for the job that needs them.
    
        
    touch_appClick a type to attach one.
    
      
    
      
    
        
    Connectors3 attached
    
        
    
        
    
          
    TypeNameScopeStatus
    
        
    
      
    
    
    
  
    

    



    
  
    
    
    
      
    Review
    
      
    Nothing lands until you say so.
    
      
    Every job ends in review. Files the expert changed in your cloud folder or repository show up as a diff, current version against proposed. Approve, reject, or send it back with feedback. The audit trail keeps every step either way.
    
    
    
    
    
      
    
        
    The diff, then the brief.
    
        
    Reviewing forty-seven files is not the point. The expert writes a brief of what it did and what it could not decide, and the diff is there when you want to check its work.
    
        
    touch_appSwitch to the brief, then approve.
    
      
    
      
    
        
    Reviewjob 8d604690 · Sort invoices
    
        
    
          
    
            
    /Finance · 3 paths · 49 files
    
            
    A2026-08/exceptions.md
    
            
    A2026-08/sorted/47 files
    
            
    DInbox/47 moved
    
          
    
          
    
            
    Finance/2026-08/exceptions.mdaddedCloud nowarrow_forwardAfter apply
    
            
     @@ -0,0 +1,14 @@
+# Invoices needing a decision · August 2026
+
+44 of 47 invoices reconciled against purchase orders in erp.
+Three did not. Nothing was posted for them.
+
+| Invoice | Supplier | Amount | Problem |
+|---|---|---|---|
+| INV-0812 | Weber Stahl | 12,480.00 € | Duplicate of INV-0790 (paid 12 Jul) |
+| INV-0819 | Hansa Packaging | 3,150.00 € | No purchase order on file |
+| INV-0833 | Nordic Fastener | 8,900.00 $ | Billed in USD against a EUR order |
+
+Suggested: reject 0812, ask purchasing for a PO on 0819,
+and hold 0833 until Nordic confirms the currency.
+
    
            
    
              
    47 invoices sorted, 3 need you
    
              
    Moved every PDF from /Finance/Inbox into /Finance/2026-08/sorted, filed by supplier and purchase order. 44 reconciled cleanly against erp.purchase_orders. Three did not, and I posted nothing for them:
    
              
    • INV-0812 duplicates INV-0790, which was paid on 12 July.
    • INV-0819 has no purchase order on file.
    • INV-0833 is billed in USD against a EUR order.
    
              
    Details in exceptions.md. Approving applies the moves and the new file to your cloud folder. Rejecting leaves the inbox untouched.
    
            
    
            
    49 changes · not appliedcheck_circleApplied to /Finance · 49 files
    
          
    
        
    
      
    
    
    

    
    
      
    
        
    Receipts
    
        
    Every job comes with a receipt.
    
        
    Tokens, cache hits, CPU and memory hours, metered per job, per user, per project. Price the same work against public cloud list prices before you decide where to run it.
    
      
    
      
    
        
    Usage & costproject Finance · 30d
    
        
    
          
    
            
    Tokens0
    
            
    Cache hit0
    
            
    vCPU-hours0
    
            
    Jobs completed0
    
          
    
          
    
            
    Tokens by model
    Three families, one ledger. Self-hosted models price through the same rate card.
    
            
    
              
      
                
    • MiniMax-M32,134,954,484
      • 
                
    • gpt-5.6-sol1,305,724,900
      • 
                
    • gemma-4-moe391,730,896
      • 
              
    
              3.8BTOKENS
            
    
          
    
          
    
            
    By category
    Every unit the ledger meters, with what it cost. Compute is your own cluster's time; tokens are what the model provider charged.
    Total €3,025.69
    
            
    CategoryUnitQuantityEventsCost
    
              
    Computegib-hour767.03399€4.60
    
              
    Computevcpu-hour383.52399€17.64
    
              
    LLMcached-prompt-token1,661,945,70242,863€107.84
    
              
    LLMcompletion-token32,676,03244,572€52.27
    
              
    LLMprompt-token2,140,683,35844,572€2,843.34
    
            
    
          
    
        
    
      
    
    
    
  
    

    



    
  
    
    
    
      
    Scale
    
      
    One worker or a thousand. Same chart.
    
      
    Workers are stateless. They pull jobs from a durable queue, hold a lease while they work, and hand the workspace back when they finish. Nothing shares memory, credentials are issued per job, and a crashed worker is just a job that gets picked up again.
    
    
    
    
    
      
    
        
    Fleetcluster main · 3 nodesAll lanes healthy
    
        
    
          
    
            
    Workers
    24pods
    
            
    1101001000
    
            
    
              
    288Jobs / hour
    
              
    €1.20Compute / hour
    
              
    0Queued
    
              
    90 sLease
    
            
    
            
      
              
    • check_circleOne job, one workspace, one set of credentials
      • 
              
    • check_circleLeases fence writers; a stolen job never answers twice
      • 
              
    • check_circleSnapshots to S3; a dead worker is a job requeued
      • 
              
    • check_circlePer-project quotas and rate cards, no surprise bill
      • 
            
    
          
    
          
    
            
    hubIngressaccount_treeOrchestrator ×2low_priorityQueue 0databasePostgres · vector · audit
    
            
    groupsWorker pool· 22 working · 2 idle
    
            
    inventory_2Workspaces 24cloud_uploadS3 snapshotsfolder_sharedNextcloud · Git · ERP
    
          
    
        
    
      
    
    
    
  
    

    



    
  
    
    
    
      
    Deploy
    
      
    Install in minutes. On anything.
    
      
    One Helm chart. A mini PC under a desk, a single node, an HA cluster, or a fleet across clusters. Postgres, vector store, audit log, identity, git, and cloud storage come with it. Bring your own if you have them.
    
    
    
    
    
      
    
        
    Installhelm · oci://ghcr.io/knaeckebrothero/charts/superhuman-remote-workerGenerate values.yaml
    
        
    
          
          
    
            
    memory
    Mini PC
    One node, one replica. Everything in one box.
    8 vCPU · 16 GiB
    
            
    dns
    Single server
    Ten workers, managed Postgres optional.
    16 vCPU · 64 GiB
    
            
    device_hub
    HA cluster
    Two orchestrators, leader-elected. Data tier replicated.
    3+ nodes
    
            
    public
    Fleet
    Workers across clusters and VM tiers, one control plane.
    any size
    
          
    
        
    
      
    
    
    
  
    

    



    
  
    
    
    
      
    Security
    
      
    Nothing leaves the building.
    
      
    SRW runs on your cluster. The only outbound connection is the model endpoint you configure, or none at all with a local model. Everything else stays where your data already lives.
    
    
    
    
    
      
      
        
    • verified_user
      Identity you already run. Keycloak OIDC with cookie sessions; the browser never holds a token. API and MCP tokens are scoped per user.
      • 
        
    • key
      Credentials per job. Connector secrets are encrypted at rest, issued to a worker only for the job that needs them, and never written into a job's config.
      • 
        
    • history
      Everything archived. Every model request and response, every tool call before and after it ran. Nothing is auto-deleted.
      • 
        
    • gavel
      Deny by default. Capability grants can only be tightened below the admin, never widened. Privileged commands wait for a named person.
      • 
        
    • hub
      Isolated workspaces. A throwaway pod or VM per job, network policies with per-project egress tiers, snapshots to your own S3.
      • 
        
    • code
      Source you can read. Fair Source: read it, patch it, run every version you install. Apache-2.0 two years after each release.
      • 
      
    
      
    
        
    What an IT lead asks first
    
        
    
          
    Does my data go to OpenAI?
    Only if you choose OpenAI. The model is picked per expert from a registry you control: a cloud API, OpenRouter, or a local endpoint.
    
          
    Who maintains this?
    A versioned Helm chart. Migrations apply at startup, an upgrade is one helm upgrade, and the source is yours to read and patch.
    
          
    What if a worker deletes something?
    It works in a throwaway workspace with its own credentials. Nothing touches your systems until you approve the diff.
    
          
    How is this different from Copilot Studio or Agentforce?
    Those are credit-billed agent builders inside a vendor's cloud. SRW is a self-hosted work queue with portable experts, a review gate, and a receipt per job.
    
          
    Isn't this CrewAI or LangGraph with a UI?
    Those are frameworks you build on. SRW is the finished system: runtime, isolation, queue, and review exist. Experts are config.
    
          
    What does a job cost?
    Tokens plus CPU and memory hours, priced against public cloud lists, shown on the receipt. No seats, no credits.
    
          
    Will workers step on each other?
    Stateless pool, durable queue with leases, one workspace per job. Scale is a replica count.
    
          
    SSO, access, audit?
    Keycloak OIDC, per-job credentials, secrets via Vault or External Secrets, every tool call and model request logged per job.
    
        
    
      
    
    
    
  
    

    


    



    
  
    
    
    
      
    Get SRW
    
      
    Free to self-host. Pilots on request.
    
      
    Run it yourself at no cost, or let us run one of your processes end to end with you and hand you the keys.
    
    
    
    
    
      
    
        
    Self-host
    
        
    Free
    
        
    Any scale, on your cluster.
    
        
      
          
    • check_circleOne Helm chart, every component included
      • 
          
    • check_circleBring your own Postgres, identity, git, S3
      • 
          
    • check_circleFair Source, Apache-2.0 after two years
      • 
          
    • check_circleCommunity support on GitHub
      • 
        
    
        Generate your values.yaml
      
    
      
    
        
    Pilot
    
        
    On request
    
        
    One process, end to end, with us on site.
    
        
      
          
    • check_circleSRW deployed in your cluster, or on a server we bring
      • 
          
    • check_circleOne process chosen with you, running with review and receipts
      • 
          
    • check_circleIntegration into your ERP, cloud storage, and identity
      • 
          
    • check_circleThe deployment, the experts, and the source stay with you
      • 
        
    
        Request a pilot
      
    
      
    
        
    Managed
    
        
    Coming soon
    
        
    The hosted edition we are building toward, run in the EU.
    
        
      
          
    • check_circleDedicated instance, your domain, your identity provider
      • 
          
    • check_circleUpdates, backups, and monitoring done for you
      • 
          
    • check_circleSame experts, connectors, and receipts
      • 
          
    • check_circleMove to self-hosting whenever you want
      • 
        
    
        Tell me when it is ready
      
    
    
    
    
    License: Functional Source License, FSL-1.1-ALv2. The only use it reserves is selling SRW itself as a hosted service. Everything you build with it is yours, and each release becomes Apache-2.0 two years after it ships.