01Implemented
Knows who is asking.
Validates the token and binds the agent to its organization. Tenant scope comes from the credential, never the request body.
- authorization
- verified
- org_id
- from token
- permissions
- enforced
The boundary between your agent and the model
Memory for agents.Control for you.
Tenant-scoped memory, bounded context, and an inspectable trace — self-hosted and OpenAI-compatible.
Inbound requestPOST/v1/chat/completions
Forwarded request.http
01Authorization: Bearer $IBEX_TOKEN
02X-IBEX-Agent-ID: agent_support
03Content-Type: application/json
04{
05"model": "gpt-4o",
06"messages": [
07{
08"role": "user",
09"content": "hello"
10}
11]
12}
data: {"choices":[{"delta":{"content":"..."}}]}
auth verifiedtenant from tokenbody unchanged
Tenant comes from the token Request lifecycle
- OpenAI-compatible
- Tenant-scoped
- Inspectable traces
01 / The boundary
One boundary. Six jobs.
Use IBEX when agents need customer context without crossing tenant lines, or when you need a bounded, inspectable path to any model.
02Implemented
Recalls only their memory.
Retrieves tenant-scoped memories from PostgreSQL and pgvector, then ranks them without letting another organization’s data into the window.
- retrieval
- pgvector
- scope
- org-filtered
- cache
- Redis hot path
03Provisional
Fits context to budget.
Calculates a token budget, packs the directive and ranked memories, and degrades to directive-only context when retrieval misses.
- budget
- computed
- ranking
- budget-aware
- fallback
- directive-only
04Implemented
Routes to the model you chose.
Keeps the OpenAI-compatible request shape while OpenAI, Anthropic, OpenAI-compatible, and mock adapters sit behind one boundary.
- openai
- native
- anthropic
- adapter
- compatible
- adapter
05Provisional
Streams without holding the client.
Forwards provider SSE, normalizes upstream errors, and keeps trace writes off the critical path. Slow-client soak evidence remains open.
- stream
- SSE forwarded
- errors
- normalized
- trace
- async write
06Provisional
Leaves evidence behind.
Records the route, token usage, and timing with a scope and date. The relay is default-off; this is local evidence, not a hosted guarantee.
- outbox
- local evidence
- relay
- default-off
- claim
- scoped
02 / Drop-in
Keep the request. Change the boundary.
Point your client at IBEX instead of the provider. The request stays OpenAI-compatible; auth, tenant scope, memory, and traces move behind the boundary.
Request lifecyclecurl -s -w "\nHTTP %{http_code}\n" \ -X POST "http://localhost:8080/v1/chat/completions" \ -H "Authorization: Bearer $IBEX_TEST_PAT" \ -H "X-IBEX-Agent-ID: $IBEX_TEST_AGENT_ID" \ -H "Content-Type: application/json" \ -d '{"model":"gpt-4o","messages":[{"role":"user","content":"hello"}]}'Tenant comes from the token. Never from the body. API reference
03 / Evidence
What works today. What we don’t claim.
IBEX is G0 and provisional. Capabilities and benchmark numbers are published with a status, a scope, and a date, so you can tell the difference between a working seam and a finished guarantee.
Local evidenceScope keptNo hosted claim
Capability status Local evidence
6 of 17 capabilities implemented locally.
13 are implemented or have a running seam. Every status is public, and every claim links back to its evidence. Hosted, HA, and production readiness are not claimed.
- Implemented6
- Provisional7
- Proposed2
- Deferred2
Local benchmark All runs
- Local proxy p99
- 1.78 ms
- HNSW recall@10
- 100.0%
- Write pipeline p95
- 32.9 ms
Not claimed: production readiness, high availability, a hosted p99 SLO, or restore drills. Local numbers keep their scope and date.
IMPLEMENTED · PROVISIONAL · PROPOSED · DEFERRED · UNKNOWN Read the full ledger
04 / Run itLocal stack · MIT
Start with mock mode
Run the boundary locally.
Start Postgres, Redis, and the services with Docker Compose. Mock mode returns a response without a provider key, so you can inspect a request before connecting a model.
- 01Prepare
Clone the repository and copy the example environment.
- 02Start the stack
Postgres, Redis, and the local services come up together.
- 03Verify the seam
Migrate, seed, and run the smoke path in mock mode.
The full quickstart covers separate service terminals and real provider keys.
git clone https://github.com/Rick1330/ibex-harness.gitcd ibex-harnesscp .env.example .envmake compose-dev-upmake db-migratemake db-seedmake dev-smokeMock mode returns 200 without a provider key. /docs/getting-started/quickstart
MIT licensed · self-hosted · no hosted control plane