Skip to content

TapPass Roadmap — H2 2026

Date: 2026-05-07 Horizon: May → December 2026 (with Q1 2027 spillover noted) Source inputs: concepts/governed-agents-architecture.md (the architectural spine), shipped state per intent-to-policy.md §10b, memory entries on OEM/UX/deployment/compliance, gap analysis vs. Giskard/Enoki.

This document is not an architecture concept. It is the operational answer to "given where we are now and what we know is missing, which big projects must we ship in the next 6 months — and in what order?"


CapabilityWhat's thereSource
Gateway (provider-agnostic)OAI + Anthropic + MCP wire formats; LiteLLM 100+ providers; capability tokens (ES256); 32-step pipeline; circuit breaker; streamingtappass/gateway/
Sandbox primitivesOpenShell with Landlock, L7 network egress, trust tiers, credential hiding via inference.localtappass/sandbox/
Audit + mandatesHash-chained audit trail; ES256-signed mandates; per-session WAL; verify_integrity passeslive on main
Intent-to-policy substrate5-layer authoring (function/categories/concerns/capabilities/pipeline), 40-line resolver, 8 catalog tables, 38 PRs merged, ~470 tests greenintent-to-policy.md §10b
Frontend (operator surface)Overview / Agents / Pipelines / Sessions / Audit Trail / Copilot / Playground / Settingsproject_frontend_scope.md
DeploymentStaging + app on Cloud Run, unified CI/CD, Terraform IaC, battle-tested pipelineproject_gcp_deployment.md
BYOK LLM keysvault_llm_keys, per-org AAD, KMS envelope, PostHog hybrid flagproject_byok_llm_keys.md
OSS / supply chainliccheck + CycloneDX SBOM gates in 4 repos; OpenChain ISO 5230 mappingproject_oss_license_compliance.md
CapabilityStateRisk
OEM decompositionOngoing per project_oem_refactor.md — protocols, container, partial migrationContends for code areas with the new keyring engine; needs sequencing
Runtime tool discoveryConcept written + 7-PR plan; capture exists in registry/tools.py (in-memory); promotion to catalog write path not shippedBlocks honest end-to-end demo per the concept's own framing
TapPass Chat (LibreChat fork)Concept written; fork not startedSMB GTM dependent; not on Collibra critical path
Async deadlock / TrustEngine migrationAsync infra ready; TrustEngine migration for tokens/credentials neededFoundational dependency for keyring derivation + sync (both touch token issuance)
UX redesignPlan written: login bg, stepped flows, context-aware agent creation, Getting Started pageParallel to spine work; not blocking
Project & Teams conceptConcept exists at projects-teams-concept.md; phase-3 frontend concept also existsRequired for org/project/team cascade in §6 of architecture

This is what the architecture document just defined and what the gap analysis surfaced:

CapabilityStatusSized
Policy compiler (with cascade merge)Concept (architecture §10); emits Compiled Policy by aspectL (~3-4 weeks)
Signed sync channel + drift detectionConcept (architecture §12)M+M (~4 weeks)
MCP-forward mode + per-org MCP registryConcept (architecture §7, §3)M+S (~3 weeks)
schema_acl + loop_guard pipeline stepsConcept (architecture §3, §6 demo)S (~1 week)
Harness + codemode + kernel exportersConcept (architecture §10 layers)S+S+M (~3 weeks)
Three CLIs (tappass, tappass-host, tappass-agent)Concept (architecture §14)M+M+S (~5 weeks)
Dashboard onboarding wizardConcept (architecture §5, §11)S (~1 week)
Pre-deployment evaluation harness + probesConcept (architecture §8)L (~3-4 weeks for MVP)
Compliance packs (EU AI Act + OWASP LLM v1)Concept (architecture §4.5)M (~3 weeks for first 2 packs)
Trust posture (SOC 2 Type 1 audit)Operational, ~6 months calendar(calendar-bound, not engineering weeks)
Public trust center expansionMarketing/opsS (~1 week of writing + legal review)

What Giskard / Enoki / the broader category have that we don't, mapped to projects we must run:

Gap (vs. competitors)Project that closes itWhy it matters
Pre-deployment red-teaming (Giskard's headline product, Enoki's "OWASP-aligned attack suites")Pre-deployment evaluation harnessProcurement gate question: "is the agent safe to ship?" — without this, every customer assembles their own answer
Vulnerability taxonomy + OWASP mappingThreat coverage taxonomy (architecture §3.2)Prepared-buyer question: "what attacks does this defend against?" — answered as content, not engineering
Regulatory compliance bundles (Giskard "EU AI Act compliance pack")Compliance packsEU/regulated buyers tick one box → compliant deployment. Differentiates against fragmented per-vendor governance
Trust posture (SOC 2, 0-training, residency)SOC 2 Type 1 audit + trust center pageFilter-question status — EU/regulated buyers won't engage architecturally without this
Drift detection in production (Enoki's third pillar)Drift detection engineCloses the loop with pre-deployment eval — production reality vs. evaluated baseline
Agent discovery in customer environmentGateway-side unenrolled-agent surfacingCustomer's first problem: "what agents do we already have?"
Public threat intelligence (Giskard's RealHarm)Public incident databaseMarketing/community asset, not architecture — but signals ongoing intelligence

Three workstreams. Critical path is the spine — without it, evaluation has nothing to run against and the Collibra demo doesn't ship.

Workstream A — Architectural spine (the moat)

Section titled “Workstream A — Architectural spine (the moat)”

These are the 12 components from architecture §14.4. Critical path for everything else.

OrderProjectSizeDependenciesQuarter
A.1TrustEngine async migration (unblocks token issuance)MQ3 (Jul)
A.2Keyring derivation engine (with cascade merge)LA.1, projects-teams substrateQ3 (Jul-Aug)
A.3Signed sync channelMA.2Q3 (Aug)
A.4tappass management CLI MVPMA.2Q3 (Aug)
A.5tappass-agent client SDKSA.3Q3 (Aug)
A.6tappass-host runtime CLI + daemonMA.3, A.5Q3 (Aug-Sep)
A.7MCP-forward mode in gatewayMA.2Q3 (Sep)
A.8schema_acl + loop_guard pipeline stepsSA.7Q3 (Sep)
A.9Per-org MCP-server registrySA.7Q3 (Sep)
A.10Kernel/sandbox profile exporterMA.6Q4 (Oct)
A.11Codemode profile exporterSA.6Q4 (Oct)
A.12Harness profile exporterSA.6Q4 (Oct)
A.13Dashboard onboarding wizardSA.4Q4 (Oct-Nov)

Total: ~13-15 engineering weeks. With 2 engineers in parallel: ~7-8 calendar weeks. With 1 engineer serial: ~13-15 weeks. A.1 is the gating dependency — must close in the first two weeks.

Workstream B — Procurement-defensibility (catching up to Giskard/Enoki)

Section titled “Workstream B — Procurement-defensibility (catching up to Giskard/Enoki)”
OrderProjectSizeDependenciesQuarter
B.1Compliance pack v1: EU AI ActMA.2 (resolver merge for packs)Q3 (Sep)
B.2Compliance pack v1: OWASP LLM Top 10MA.2Q3 (Sep)
B.3Pre-deployment evaluation harness MVPLA.5 (uses tappass-agent SDK)Q4 (Oct-Nov)
B.4Probe library v1 (OWASP LLM + EU AI Act)MB.3Q4 (Nov)
B.5Drift detection engineMA.3 (sync gives us baselines)Q4 (Nov-Dec)
B.6Compliance packs v2: GDPR, PCI-DSS, HIPAAMB.1 (template established)Q1 2027
B.7Compliance packs v3: NIS2, DORAMB.1Q1 2027

Total: ~10 engineering weeks for v1 (B.1-B.5). B.6 and B.7 are demand-driven.

Workstream C — Trust + concrete artifacts

Section titled “Workstream C — Trust + concrete artifacts”
OrderProjectSizeDependenciesQuarter
C.1Trust center page expansion (0-training, residency, encryption, sub-processors)S— (independent)Q3 (Jul)
C.2SOC 2 Type 1 readiness assessment + auditor selection(operations)C.1Q3 (Jul-Aug)
C.3SOC 2 Type 1 audit fieldwork(operations)C.2Q3-Q4 (Sep-Dec)
C.4SOC 2 Type 1 report issued(operations)C.3Q4 (Dec)
C.5tappass/collibra-agent reference repo scaffoldedMA.7 (MCP proxy), A.13 (wizard), B.3 (eval)Q4 (Nov-Dec)
C.6Collibra demo to Stijn / Nick (Spoor 1 close)(event)C.5Q4 (Dec)

SOC 2 is calendar-bound — engineering doesn't accelerate it past ~6 months. Start in July, certificate by December.

Workstream D — Pre-existing flight that must continue

Section titled “Workstream D — Pre-existing flight that must continue”

These are not new projects; they are obligations that must close to unblock the spine.

ProjectStatusWhy it matters for H2
OEM decompositionIn progressContends for the same code areas as keyring engine. Must finish or coordinate carefully with A.2.
Runtime tool discovery (capture → catalog promotion)7-PR plan pendingBlocks the honest end-to-end demo per the concept's own framing. Should ship before C.5.
Project & Teams substrateConcept written; phase-3 frontend existsRequired for cascade in A.2. Must ship by August.
UX redesignPlan writtenParallel to spine; not blocking; touches A.13 wizard look-and-feel.

Workstream E — Deferred / explicit non-priorities

Section titled “Workstream E — Deferred / explicit non-priorities”

These exist as concepts/initiatives but should NOT take H2 2026 cycles:

  • TapPass Chat (LibreChat fork) — SMB GTM bet; only ship if SMB pipeline materializes. Otherwise H1 2027.
  • Compliance packs v2/v3 — demand-driven; ship when first customer asks for them.
  • Public threat intelligence asset (RealHarm-equivalent) — marketing project; H1 2027.
  • Subsequent reference agents (support emailer, code reviewer, refund processor) — only after collibra-agent proves the architecture. Q1 2027.
  • Agent discovery (full mechanism beyond passive surfacing) — the architecture commits to surfacing the signal; the full discovery + auto-classification flow is a separate concept.

│ Jul │ Aug │ Sep │ Oct │ Nov │ Dec │
─────────────────────┼─────┼─────┼─────┼─────┼─────┼─────┤
A.1 TrustEngine │ ███ │ │ │ │ │ │
A.2 Keyring deriv. │ █ │ ███ │ ██ │ │ │ │
A.3 Sync channel │ │ ███ │ │ │ │ │
A.4 tappass CLI │ │ ███ │ │ │ │ │
A.5 tappass-agent │ │ ██ │ │ │ │ │
A.6 tappass-host │ │ ██ │ ███ │ │ │ │
A.7 MCP-forward │ │ │ ███ │ │ │ │
A.8 ACL+loop guard │ │ │ ██ │ │ │ │
A.9 MCP registry │ │ │ █ │ │ │ │
A.10 Kernel exporter │ │ │ │ ██ │ │ │
A.11 Codemode export │ │ │ │ ██ │ │ │
A.12 Harness export │ │ │ │ ██ │ │ │
A.13 Wizard │ │ │ │ █ │ ██ │ │
B.1 EU AI Act pack │ │ │ ███ │ │ │ │
B.2 OWASP pack │ │ │ ███ │ │ │ │
B.3 Eval harness │ │ │ │ ███ │ ██ │ │
B.4 Probe library v1 │ │ │ │ │ ██ │ │
B.5 Drift engine │ │ │ │ │ ███ │ ██ │
C.1 Trust center │ ██ │ │ │ │ │ │
C.2-C.4 SOC 2 │ │ ███ │ ███ │ ███ │ ███ │ ██ │
C.5 collibra-agent │ │ │ │ │ ███ │ ██ │
C.6 Demo Collibra │ │ │ │ │ │ █ │
─────────────────────┴─────┴─────┴─────┴─────┴─────┴─────┘
▲ ▲
Spine MVP ready Collibra demo

Critical path: A.1 → A.2 → A.3+A.4+A.5+A.6 → A.7+A.8 → C.5 → C.6. Roughly 6 calendar months end-to-end with 2 engineers; 8-9 months with 1.


Assuming 2 platform engineers + 1 frontend + 1 part-time compliance-content owner:

WorkstreamOwner profileEstimated weeks (calendar)
A (spine)2 platform eng13-15 weeks parallel work
B (procurement)1 platform eng + part-time content10 weeks (B.1-B.5)
C (trust + reference)Operations + 1 platform eng for C.56 months calendar (mostly external)
D (in-flight obligations)Existing ownersvaries; assume 30% drag on platform eng capacity

If headcount is tighter (1 platform engineer), the realistic shape becomes:

  • Q3: A.1 → A.2 → A.3 only
  • Q4: A.4-A.6 + B.1+B.2 (compliance content) + C.1 (trust page)
  • Q1 2027: A.7-A.13 + B.3-B.5 + C.5 (Collibra demo)

In which case Collibra demo slips to Feb-Mar 2027.


6. Strategic prioritization — what to fight for

Section titled “6. Strategic prioritization — what to fight for”

Non-negotiable for the year:

  • A.1-A.6 (spine MVP): without this, nothing else has a substrate. Must close by end Q3.
  • B.1-B.2 (compliance packs v1): with EU AI Act + OWASP LLM packs in the bag, we have a procurement-defensible answer that Giskard's parity-claim doesn't dominate.
  • C.1-C.4 (SOC 2 Type 1): filter-question status. Without it, EU/regulated conversations don't start.
  • C.5-C.6 (Collibra demo): the partner conversation is the forcing function. December close is realistic if A.1-A.6 hit Q3.

Worth fighting for:

  • B.3-B.5 (eval + drift): closes the Giskard/Enoki competitive gap. December delivery makes us competitive against either of them in a head-to-head.
  • A.7-A.13 (full enforcement layer + wizard): brings the full architecture to the demo. Without it, the demo is "policy enforcement" but not "every layer."

Not worth fighting for in 2026:

  • TapPass Chat — only if SMB pipeline materializes.
  • Compliance packs v2/v3 — demand-driven.
  • Subsequent reference agents — after Collibra ships.
  • Public threat intelligence — marketing project; can wait.

  1. TrustEngine async migration (A.1) takes longer than 2 weeks. Foundational; everything blocks. Mitigation: scope explicitly in week 1, accept narrow MVP.
  2. OEM decomposition contention. Same code areas as keyring engine. Mitigation: weekly sync between OEM owner and keyring lead; merge windows clearly defined.
  3. SOC 2 audit slippage. Auditor availability is the main calendar risk. Mitigation: select auditor in July, lock fieldwork dates in August.
  4. Collibra partnership cools. Demo loses urgency. Mitigation: ship the spine and reference repo regardless — collibra-agent is the first instance of an architecture that supports many; the work isn't wasted.
  5. Compliance pack content correctness. If EU AI Act mapping is wrong, we lose the procurement-defensibility we'd be claiming. Mitigation: external legal review before claiming any pack publicly.
  6. Pre-deployment eval scope creep. Easy to overinvest in probe coverage. Mitigation: ship MVP with OWASP LLM Top 10 only; expand demand-driven.

To make this roadmap real, the following decisions are blocking:

DecisionOwnerLean
Headcount commitment for A and B workstreamsFounders2 platform eng + 1 frontend at minimum
TrustEngine async migration scopePlatform leadNarrow MVP unblocking keyring — defer broader migration
SOC 2 auditor selectionFounders / opsEngage 2-3 firms in next 2 weeks; pick by end of June
Compliance content ownershipFoundersHire / contract a Compliance / GRC owner for B workstream
OEM decomposition exit criteriaExisting OEM ownerDefine what "done" looks like and target date
Collibra demo target dateFounders + Stijn/NickLock to a December week to backlog from

By end of December 2026, TapPass should have:

  • Compiled Policy compiler + signed sync shipped end-to-end
  • Three CLIs (tappass, tappass-host, tappass-agent) usable
  • EU AI Act + OWASP LLM Top 10 compliance packs live
  • Pre-deployment evaluation harness MVP with 50+ probes
  • SOC 2 Type 1 report in hand
  • tappass/collibra-agent reference repo demonstrated to Collibra technical leadership
  • Trust center page covering 0-training, residency, encryption, certifications
  • At least one external customer using a non-Collibra reference agent in evaluation

That set of artifacts puts TapPass at parity with Giskard on procurement-defensibility and at depth-of-enforcement no other vendor offers — the unique combination that this architecture concept argues for.


10. Where this roadmap lives and how it stays current

Section titled “10. Where this roadmap lives and how it stays current”

This document is operational. It should be:

  • Reviewed monthly against actual delivery
  • Updated when scope, sequencing, or headcount changes
  • Used as the input to quarterly planning conversations
  • Anchored to the architecture concept (concepts/governed-agents-architecture.md) — when the architecture changes, this roadmap re-derives

It is not a contract — it is a planning artifact. Specs, PRs, and ticket trackers are downstream of this.