CI/CD for agent intelligence

Ship everything
your agents learn.

Agentevo gives coding agents a lifecycle for experience: successful runs become reusable skills, failures and corrections become regression tests and guards, and proven skills compile into deterministic workflows. Raw data stays local by default.

Start locally See how it works

Free tier · 10,000 cloud requests/mo · No card · Works offline

quickstart — zsh 
$ pip install agentevo-sdk
$ agentevo capture ./session.jsonl --adapter claude-code
✓ session 7f3a… normalized: 214 events
$ agentevo recall "fix flaky test" --max-objects 3
✓ 3 skills · compat EXACT · risk ≤ medium
Local guard p95
<50ms
Cloud recall p95
<250ms
Raw traces to cloud
Zero
Works with your agent
  • Claude Code
  • Codex
  • Cursor
  • OpenCode
  • CI / QA agents

01

The learning loop.

Every run makes the next one better — or safer. Select a stage to inspect it.

    02

    Set up in three steps.

    One runtime on your machine. No pipeline to build, no data to upload.

    Capture every run, locally

    Adapters normalize provider events into one canonical contract — event_id, session_id, monotonic seq, typed payload, content hash. DLP strips secrets before anything is persisted.

    • Raw prompts, diffs and logs stay in your encrypted local vault
    • Duplicate event_id values are ignored deterministically
    • Capture keeps working fully offline
    capture
    $ agentevo capture ./session.jsonl --adapter claude-code
✓ session 7f3a… normalized: 214 events
✓ DLP: 3 secrets redacted pre-persistence
✓ vault: encrypted (AES-256, OS keychain)

    03

    Trust is earned.

    A signature proves authenticity — never trust. Trust comes only from verification evidence plus policy-authorized promotion.

      Lateral states: Quarantined Stale Rejected Revoked — none of them auto-execute. Revocation writes a tombstone, never a delete.

      04

      Local-first, provably.

      Classification is the architecture. There is no cloud raw-trace endpoint to misconfigure.

      Data classification

      ClassExamplesRule
      C0SDK versions, public schemasAnywhere
      C1Tenant IDs, plan, countersOnly default cloud payload
      C2Sanitized steps, fingerprintsOpt-in sync, tenant-scoped, signed
      C3Raw prompts, code, logsLocal / VPC only
      C4Secrets, tokens, keysNever a valid payload — anywhere

      Privacy modes

      • Local-only default
        Cloud sees control metadata (C1) and nothing else.
      • Opaque Relay
        Cloud stores E2EE envelopes; clients index and decrypt locally.
      • Searchable Managed Sync
        Sanitized C2 with tenant-scoped server recall.
      • Customer VPC
        Same contract, your keys, your indexes, your boundary.

      Hard guarantees

      • Retrieved memory is never privileged instruction
      • Tenant and workspace derive from the authenticated principal
      • High-risk auto-execution fails closed on stale policy
      • Local capture, guard and recall work offline and at quota

      05

      Start free. Scale sanely.

      One batched call counts as one request. Local work, usage reads and job polling are never billed.

      Free

      $0/mo

      For solo builders

      • 10,000 cloud requests / month
      • Local runtime, unlimited
      • 1 user · 1 workspace
      • Never overages — local keeps working
      Start free

      Most popular

      Pro

      $49/mo

      For professionals

      • 100,000 cloud requests / month
      • Managed sync + cloud recall
      • Async verify orchestration
      • BYOK inference
      Get early access

      Team

      $299/mo

      For fleets

      • 1,000,000 cloud requests / month
      • Shared registry + RBAC + audit
      • Team analytics
      • 10 seats included
      Talk to us

      Enterprise

      Custom

      For regulated orgs

      • Customer VPC / on-prem
      • SSO, SCIM, CMK, residency
      • SLA + private connectors
      • Security review support
      Book a pilot

      Pro, Team and Enterprise limits are launch targets and may change during early access. Free — 10,000 requests — is a product guarantee.

      06

      Questions, answered.

      Where does my agent's raw output go?

      Nowhere, by default. Raw prompts, tool I/O, code and logs (C3) live in your encrypted local vault. The cloud control plane receives control metadata (C1) only. Sanitized derived data (C2) syncs only in an explicit sync mode, tenant-scoped and signed. There is no cloud raw-trace endpoint.

      Which agents work with Agentevo?

      Any agent with an adapter. We ship Claude Code and Codex today, with Cursor and OpenCode following — plus CI, QA and security agents. Adapters map provider events into one canonical contract and must pass conformance fixtures; a connector update can never silently elevate classification or permissions.

      What latency does this add to my runs?

      Local guard evaluation targets p95 under 50ms and runs offline. Cloud recall targets p95 under 250ms. Extraction, compilation and verification are async jobs — they never block the agent loop.

      Do I need Agentevo's models? What is BYOK?

      No. Model inference is bring-your-own-key by default: your agent keeps using its own provider and keys. A managed inference add-on is available separately. Note the honest boundary: local-first governs data sent to Agentevo — your agent's own calls to its model provider are disclosed separately.

      What happens when I hit my quota?

      On Free, cloud value operations pause and everything local — capture, guard, recall, registry — keeps working. Free never overages. Paid plans can auto-overage or hard-cap per admin setting, with an 80% warning first.

      How do skills stay safe as my repo changes?

      Every object carries a layered environment fingerprint. When dependencies, tools, models, schemas or policy drift, compatibility is reclassified and stale skills stop auto-executing instead of running blind. Revocations propagate as tombstones and take priority over ordinary sync.

      Give your agents a memory worth trusting.

      Install locally, run your agent, and watch the first verified skill land. Your data never leaves the machine unless you say so.

      Request early access support@agentevo.sh
      get started
      $ pip install agentevo-sdk
# local-only is the default — nothing leaves your machine
$ agentevo capture ./session.jsonl --adapter claude-code
$ agentevo recall "deploy checklist" --max-objects 5