One model provider in front of every AI account you own — routing is automatic, safety is first.
AgentPool is a small, local-first inference service. Your agent runtime (OpenCode first) sees a
single model provider called agentpool. Behind it sits a pool of independently authenticated
routes — API keys, subscription-backed accounts, enterprise endpoints and local models — and for
every new inference request AgentPool automatically picks an eligible, certified route based on the
requested model alias, capabilities, permissions, session continuity and remaining capacity.
When one route runs out of capacity, cools down or becomes ineligible, the next request moves to another without anyone pressing a button. In-flight streams stay pinned to the route they started on. Nothing is silently swapped for a different model.
Host (e.g. TradeAgent) ── controls agents ──▶ OpenCode ── model requests ──▶ AgentPool ──▶ route A │ B │ C …
(one provider) (your accounts)
TradeAgent owns the work. OpenCode executes the agents. AgentPool supplies the inference. AgentPool never owns an agent, its task, its session, its tools, its files or its memory.
The design (spec v5.2) puts correctness and safety ahead of features:
- Provider only. AgentPool is a model endpoint, never an agent runner, task queue or tool executor.
- Fail closed. Unknown capabilities, unverified routes and ambiguous retries are refused, not guessed.
- No silent substitution. A request for model X is never quietly answered by unrelated model Y.
- Honest status.
UNKNOWNis never treated asPROVEN; nothing is called working until a pinned, real-client conformance test passes. - Credentials stay put. Upstream keys live only in the AgentPool process and the OS vault — never in OpenCode's config, prompts or a worker's environment.
- Loopback only. No LAN/WAN listener in v1; browser-origin requests are rejected.
- AI-assisted, never AI-governed. An optional route-risk advisor (JEV) may suggest soft ranking penalties off the hot path; deterministic code alone decides capacity, policy, billing, credentials and retries.
v0.1.0 — skeleton. The binary builds, serves /healthz on loopback, rejects browser origins,
and answers every inference path with an explicit 501 not implemented. No routing, credentials or
upstream calls exist yet. See ROADMAP.md for milestones P0–P7 and their gates.
make test # vet + race-enabled tests
make build # ./bin/agentpool
./bin/agentpool serve
curl http://127.0.0.1:43819/healthzRequires Go 1.26 or newer.
AgentPool pools accounts that one owner holds, for that owner's own use. It is not a tool for sharing or reselling accounts between people. A connector ships only if it uses a flow its provider documents and whose terms permit that use — see docs/PROVIDER_POLICY.md.
| Document | Purpose |
|---|---|
| Spec v5.2 (current) | Normative implementation specification |
| docs/architecture/ | Spec index, version history, module map |
| docs/PROVIDER_POLICY.md | Which connectors may exist and how they are certified |
| examples/opencode.jsonc | Illustrative OpenCode provider config (unverified) |
| ROADMAP.md | Milestones and release gates |
Software releases follow Semantic Versioning via git tags (v0.1.0), with
changes recorded in CHANGELOG.md. The architecture version (currently 5.2.0) is
separate and lives in the spec's header. Before 1.0.0, minor releases may change interfaces.
See CONTRIBUTING.md. Security issues: SECURITY.md.
GNU Affero General Public License v3.0. If you run a modified AgentPool as a network service, you must offer its source to that service's users.
Built by Nicolas Beeckman.
The Go skeleton, tests, CI and repository documentation were written with Claude Opus 5.5 (Anthropic), via Claude Code, working from the v5.2 specification.