Keep your model. Keep your harness. Fix everything around them.

Condensate attaches to the harness you already run through two thin seams — one MCP endpoint and one hook — and does the surrounding work: capabilities, identity, steering, honest measurement. Your model traffic goes straight to your provider, untouched.

Get started Default mode is observe

First resident suite, live today

session 4c9f · hook events · consent: steer

PostToolUse  Grep {"pattern": "registerRoute"}                      ✓
PostToolUse  Grep {"pattern": "registerRoute"}    signature repeat x2
PostToolUse  Grep {"pattern": "registerRoute"}    signature repeat x3
tool_loop fired  threshold 3 → template context_note_v1
inject "tool loop detected: Grep has been called 3 times in a row
        with identical arguments... vary the arguments."
PostToolUse  Read src/router.ts       signature changed — loop broken
ledger #4821  fire recorded · digest only, no transcript

2seams: one endpoint, one hook

0touches on your model traffic

~45msfail-open when the backend is down

500session floor before any eval claim

How it attaches

Exactly two surfaces.

The client contract is deliberately dumb and deliberately stable. Everything intelligent lives server-side — you never reinstall anything.

01 One MCP endpoint

Add a single server; the whole suite fans in server-side — namespaced, filtered to your principal. Adding a capability never touches your config again.

02 One hook shim

A dependency-free binary your harness's own hook mechanism invokes. Events out, verdicts in. Backend down? It gets out of the way in ~45ms — your agent never depends on condensate being up.

config

// one server, every capability — fanned in server-side
{
  "mcpServers": {
    "condensate": {
      "type": "http",
      "url": "https://<backend>/mcp",
      "headers": { "x-api-key": "<executor key>" }
    }
  }
}

// your harness's own hook mechanism invokes the shim
{
  "hooks": {
    "PreToolUse":      [{ "hooks": [{ "type": "command", "command": "condensate-hook" }] }],
    "PostToolUse":     [{ "hooks": [{ "type": "command", "command": "condensate-hook" }] }],
    "UserPromptSubmit": [{ "hooks": [{ "type": "command", "command": "condensate-hook" }] }],
    "Stop":            [{ "hooks": [{ "type": "command", "command": "condensate-hook" }] }]
  }
}

# default mode is observe — nothing intervenes without opt-in.
# watch the ledger before you let it steer:
# every detector fire AND every suppression is recorded,
# append-only, digests not transcripts.

Attaches through the hook mechanism your harness already has. Claude Code today; more adapters landing. Contract: hook-contract.md.

Steering

Your agent loops, stalls, and quits early. Something should notice.

Detectors are pure state machines over hook events. They emit only booleans and scores; the composer speaks only from a fixed template library — there is no path from task content to injected text. Consent defaults to observe; nothing intervenes without opt-in; every fire and every suppression is ledgered.

tool_loop detector

Same tool, identical canonical args, N times in a row. Signature compare, nothing read.

edit_thrash detector

The same file rewritten back and forth without the tests moving.

progress_stall detector

Activity without state change — busy signals, no delta.

premature_stop detector

Stop event while the session's own task state still shows open work.

filler_inflation detector

Output growing while information doesn't.

composer templates only

Verdicts carry a template_id plus scalar params. No free-form generation, ever.

the eval layer refuses to flatter honesty

sha-bucketed holdout arm runs alongside every steered session
no outcome claims below 500 pooled sessions — the floor has not been hit, so no numbers are on this page

The suite

Onchain is the first resident, not the product.

live today

The first resident: onchain. 22 EVM tools over MCP — reads, traces, logs, ABIs, verified source, address labels, cross-chain status. BYOK and self-host free, or metered per call from a wallet via x402. Docs: byok.md · x402-gateway.md · endpoint at onchain.condensate.dev/mcp.

The generalization: executor-held connections mean any integration you register rides the same endpoint, same auth, same metering. Onchain proves the plane.

Trust

never your model traffic

Routing is advisory only. Requests terminate at your own subscription endpoints under your own auth. Condensate is not in the request path — by design decision, not configuration.

content-free by construction

Prompt text never persists and never reaches inference. Detectors keep digests, not transcripts. Steering text comes from templates, never from your content.

fail-open

Backend unreachable means your agent runs exactly as it did before condensate existed. The shim's job is to get out of the way in ~45ms.

Identity is bought (WorkOS); custody is owned (executor.sh, MIT, self-hosted). Buy identity, own custody.

What ships today vs in flight

component	status	proof
hook shim + contract v1	shipped	merged; fail-open ~45ms verified
detectors + consent + ledger	shipped	merged: detectors, consent tiers, append-only ledger, checkpoints, holdout, eval harness
identity + executor connections	shipped	WorkOS merged, verified live against staging; full local emulation with zero external accounts
model-routing + executor-auth	shipped	work-kind→model table extracted byte-faithful; fail-closed key validation
onchain MCP suite	shipped	live: gateway, x402 metering, BYOK, self-host
harness-agnostic core + adapters	landing	proven in review, merging now; Claude Code byte-compat proven
/mcp aggregation gateway	landing	proven in review, merging now; live-proven: 30 tools, upstream-failure isolation

This board mirrors the architecture doc's status board. The site never states more than it does.

Get started

consent first

Add the endpoint, install the hook, and stop there. Default mode is observe. Watch the ledger before you let it steer. Config for both surfaces is in how it attaches above.