TapPass Roadmap — H2 2026
TapPass Roadmap — H2 2026
Section titled “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?"
1. Where we are now (the honest snapshot)
Section titled “1. Where we are now (the honest snapshot)”1.1 Shipped and mature
Section titled “1.1 Shipped and mature”| Capability | What's there | Source |
|---|---|---|
| Gateway (provider-agnostic) | OAI + Anthropic + MCP wire formats; LiteLLM 100+ providers; capability tokens (ES256); 32-step pipeline; circuit breaker; streaming | tappass/gateway/ |
| Sandbox primitives | OpenShell with Landlock, L7 network egress, trust tiers, credential hiding via inference.local | tappass/sandbox/ |
| Audit + mandates | Hash-chained audit trail; ES256-signed mandates; per-session WAL; verify_integrity passes | live on main |
| Intent-to-policy substrate | 5-layer authoring (function/categories/concerns/capabilities/pipeline), 40-line resolver, 8 catalog tables, 38 PRs merged, ~470 tests green | intent-to-policy.md §10b |
| Frontend (operator surface) | Overview / Agents / Pipelines / Sessions / Audit Trail / Copilot / Playground / Settings | project_frontend_scope.md |
| Deployment | Staging + app on Cloud Run, unified CI/CD, Terraform IaC, battle-tested pipeline | project_gcp_deployment.md |
| BYOK LLM keys | vault_llm_keys, per-org AAD, KMS envelope, PostHog hybrid flag | project_byok_llm_keys.md |
| OSS / supply chain | liccheck + CycloneDX SBOM gates in 4 repos; OpenChain ISO 5230 mapping | project_oss_license_compliance.md |
1.2 In flight / partial
Section titled “1.2 In flight / partial”| Capability | State | Risk |
|---|---|---|
| OEM decomposition | Ongoing per project_oem_refactor.md — protocols, container, partial migration | Contends for code areas with the new keyring engine; needs sequencing |
| Runtime tool discovery | Concept written + 7-PR plan; capture exists in registry/tools.py (in-memory); promotion to catalog write path not shipped | Blocks honest end-to-end demo per the concept's own framing |
| TapPass Chat (LibreChat fork) | Concept written; fork not started | SMB GTM dependent; not on Collibra critical path |
| Async deadlock / TrustEngine migration | Async infra ready; TrustEngine migration for tokens/credentials needed | Foundational dependency for keyring derivation + sync (both touch token issuance) |
| UX redesign | Plan written: login bg, stepped flows, context-aware agent creation, Getting Started page | Parallel to spine work; not blocking |
| Project & Teams concept | Concept exists at projects-teams-concept.md; phase-3 frontend concept also exists | Required for org/project/team cascade in §6 of architecture |
1.3 Concept-stage / not started
Section titled “1.3 Concept-stage / not started”This is what the architecture document just defined and what the gap analysis surfaced:
| Capability | Status | Sized |
|---|---|---|
| Policy compiler (with cascade merge) | Concept (architecture §10); emits Compiled Policy by aspect | L (~3-4 weeks) |
| Signed sync channel + drift detection | Concept (architecture §12) | M+M (~4 weeks) |
| MCP-forward mode + per-org MCP registry | Concept (architecture §7, §3) | M+S (~3 weeks) |
schema_acl + loop_guard pipeline steps | Concept (architecture §3, §6 demo) | S (~1 week) |
| Harness + codemode + kernel exporters | Concept (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 wizard | Concept (architecture §5, §11) | S (~1 week) |
| Pre-deployment evaluation harness + probes | Concept (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 expansion | Marketing/ops | S (~1 week of writing + legal review) |
2. The gap → project map
Section titled “2. The gap → project map”What Giskard / Enoki / the broader category have that we don't, mapped to projects we must run:
| Gap (vs. competitors) | Project that closes it | Why it matters |
|---|---|---|
| Pre-deployment red-teaming (Giskard's headline product, Enoki's "OWASP-aligned attack suites") | Pre-deployment evaluation harness | Procurement gate question: "is the agent safe to ship?" — without this, every customer assembles their own answer |
| Vulnerability taxonomy + OWASP mapping | Threat 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 packs | EU/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 page | Filter-question status — EU/regulated buyers won't engage architecturally without this |
| Drift detection in production (Enoki's third pillar) | Drift detection engine | Closes the loop with pre-deployment eval — production reality vs. evaluated baseline |
| Agent discovery in customer environment | Gateway-side unenrolled-agent surfacing | Customer's first problem: "what agents do we already have?" |
| Public threat intelligence (Giskard's RealHarm) | Public incident database | Marketing/community asset, not architecture — but signals ongoing intelligence |
3. Prioritized projects for H2 2026
Section titled “3. Prioritized projects for H2 2026”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.
| Order | Project | Size | Dependencies | Quarter |
|---|---|---|---|---|
| A.1 | TrustEngine async migration (unblocks token issuance) | M | — | Q3 (Jul) |
| A.2 | Keyring derivation engine (with cascade merge) | L | A.1, projects-teams substrate | Q3 (Jul-Aug) |
| A.3 | Signed sync channel | M | A.2 | Q3 (Aug) |
| A.4 | tappass management CLI MVP | M | A.2 | Q3 (Aug) |
| A.5 | tappass-agent client SDK | S | A.3 | Q3 (Aug) |
| A.6 | tappass-host runtime CLI + daemon | M | A.3, A.5 | Q3 (Aug-Sep) |
| A.7 | MCP-forward mode in gateway | M | A.2 | Q3 (Sep) |
| A.8 | schema_acl + loop_guard pipeline steps | S | A.7 | Q3 (Sep) |
| A.9 | Per-org MCP-server registry | S | A.7 | Q3 (Sep) |
| A.10 | Kernel/sandbox profile exporter | M | A.6 | Q4 (Oct) |
| A.11 | Codemode profile exporter | S | A.6 | Q4 (Oct) |
| A.12 | Harness profile exporter | S | A.6 | Q4 (Oct) |
| A.13 | Dashboard onboarding wizard | S | A.4 | Q4 (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)”| Order | Project | Size | Dependencies | Quarter |
|---|---|---|---|---|
| B.1 | Compliance pack v1: EU AI Act | M | A.2 (resolver merge for packs) | Q3 (Sep) |
| B.2 | Compliance pack v1: OWASP LLM Top 10 | M | A.2 | Q3 (Sep) |
| B.3 | Pre-deployment evaluation harness MVP | L | A.5 (uses tappass-agent SDK) | Q4 (Oct-Nov) |
| B.4 | Probe library v1 (OWASP LLM + EU AI Act) | M | B.3 | Q4 (Nov) |
| B.5 | Drift detection engine | M | A.3 (sync gives us baselines) | Q4 (Nov-Dec) |
| B.6 | Compliance packs v2: GDPR, PCI-DSS, HIPAA | M | B.1 (template established) | Q1 2027 |
| B.7 | Compliance packs v3: NIS2, DORA | M | B.1 | Q1 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”| Order | Project | Size | Dependencies | Quarter |
|---|---|---|---|---|
| C.1 | Trust center page expansion (0-training, residency, encryption, sub-processors) | S | — (independent) | Q3 (Jul) |
| C.2 | SOC 2 Type 1 readiness assessment + auditor selection | (operations) | C.1 | Q3 (Jul-Aug) |
| C.3 | SOC 2 Type 1 audit fieldwork | (operations) | C.2 | Q3-Q4 (Sep-Dec) |
| C.4 | SOC 2 Type 1 report issued | (operations) | C.3 | Q4 (Dec) |
| C.5 | tappass/collibra-agent reference repo scaffolded | M | A.7 (MCP proxy), A.13 (wizard), B.3 (eval) | Q4 (Nov-Dec) |
| C.6 | Collibra demo to Stijn / Nick (Spoor 1 close) | (event) | C.5 | Q4 (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.
| Project | Status | Why it matters for H2 |
|---|---|---|
| OEM decomposition | In progress | Contends 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 pending | Blocks the honest end-to-end demo per the concept's own framing. Should ship before C.5. |
| Project & Teams substrate | Concept written; phase-3 frontend exists | Required for cascade in A.2. Must ship by August. |
| UX redesign | Plan written | Parallel 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-agentproves 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.
4. Timeline at a glance
Section titled “4. Timeline at a glance” │ 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 demoCritical 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.
5. Resource implications (rough)
Section titled “5. Resource implications (rough)”Assuming 2 platform engineers + 1 frontend + 1 part-time compliance-content owner:
| Workstream | Owner profile | Estimated weeks (calendar) |
|---|---|---|
| A (spine) | 2 platform eng | 13-15 weeks parallel work |
| B (procurement) | 1 platform eng + part-time content | 10 weeks (B.1-B.5) |
| C (trust + reference) | Operations + 1 platform eng for C.5 | 6 months calendar (mostly external) |
| D (in-flight obligations) | Existing owners | varies; 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.
7. Top risks
Section titled “7. Top risks”- TrustEngine async migration (A.1) takes longer than 2 weeks. Foundational; everything blocks. Mitigation: scope explicitly in week 1, accept narrow MVP.
- OEM decomposition contention. Same code areas as keyring engine. Mitigation: weekly sync between OEM owner and keyring lead; merge windows clearly defined.
- SOC 2 audit slippage. Auditor availability is the main calendar risk. Mitigation: select auditor in July, lock fieldwork dates in August.
- Collibra partnership cools. Demo loses urgency. Mitigation: ship the spine and reference repo regardless —
collibra-agentis the first instance of an architecture that supports many; the work isn't wasted. - 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.
- Pre-deployment eval scope creep. Easy to overinvest in probe coverage. Mitigation: ship MVP with OWASP LLM Top 10 only; expand demand-driven.
8. What to decide in the next 2 weeks
Section titled “8. What to decide in the next 2 weeks”To make this roadmap real, the following decisions are blocking:
| Decision | Owner | Lean |
|---|---|---|
| Headcount commitment for A and B workstreams | Founders | 2 platform eng + 1 frontend at minimum |
| TrustEngine async migration scope | Platform lead | Narrow MVP unblocking keyring — defer broader migration |
| SOC 2 auditor selection | Founders / ops | Engage 2-3 firms in next 2 weeks; pick by end of June |
| Compliance content ownership | Founders | Hire / contract a Compliance / GRC owner for B workstream |
| OEM decomposition exit criteria | Existing OEM owner | Define what "done" looks like and target date |
| Collibra demo target date | Founders + Stijn/Nick | Lock to a December week to backlog from |
9. Six-month outcome we're optimizing for
Section titled “9. Six-month outcome we're optimizing for”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-agentreference 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.