↯ SwiftAgent

LOCAL AGENT CONTROL ROOM

Your workspace.
Your choice of agent.

One local place to run, inspect, approve, resume, and compare Claude Code, Codex, OpenCode, ACP agents, or a restricted command adapter—without pretending their capabilities are identical.

Claude CodeCodexOpenCodeACP v1+ command
Run it locally Inspect the code
01
Local-first
02
Agent-aware
03
Evidence-led
NEW RUNLOCAL
C

Codexapp-server v2 · ready

⌄

Fix the release badge and run its acceptance test. Keep changes inside the demo workspace.

Draft stays when the agent changes
✓ approvals✓ plans✓ usageworkspace-write
demo-workspace/runs/codexRun →

Controls come from the selected adapter’s declared capabilities.

01 / THE POINT

Change the agent.
Keep the operating model.

SwiftAgent normalizes the workflow around a run—not the agent’s personality. Each adapter keeps its native protocol, authentication, session ID, model, and safety posture. The interface adds one durable timeline, one receipt format, and a deliberate way to hand reviewed context to another agent.

02 / CHOOSE AN ENTRY POINT

Agent-specific details.
Agent-neutral operating model.

stream-json

Claude Code

Native session resume and tool evidence, with explicit unsupported controls.

Explore Claude Code → app-server v2

Codex

Approvals, questions, plans, usage, and separate native safety evidence.

Explore Codex → ACP first

OpenCode

Negotiated structured features with a visibly reduced JSON fallback.

Explore OpenCode → Adapter API 1.0

Any ACP agent

A strict local manifest, public contract kit, and no automatic download.

Connect an ACP agent →

03 / ONE FOCUSED LOOP

Choose. Run. Review. Handoff.

  1. 01

    Detect locally

    See installed agents, versions, models, auth state, protocol, and readiness without making a model call.

  2. 02

    Write once

    Your draft and workspace stay put while you compare ready agents and capability-aware controls.

  3. 03

    Review evidence

    Tool events, approvals, questions, Git impact, safety layers, and verification live in one Local Run Receipt.

  4. 04

    Handoff deliberately

    Preview and redact bounded context before a different agent starts a new run. Native sessions never cross.

04 / LOCAL RUN RECEIPT

A shared record,
not a flattened story.

Every run stores common evidence and expandable native details. SwiftAgent records test status only when explicit evidence exists; an assistant saying “done” is not a passing test.

  • Agent, adapter, protocol, model, and native session
  • Normalized timeline with native event identity
  • Git baseline, changed files, and bounded diff summary
  • Native safety and SwiftAgent isolation shown separately
LOCAL RUN RECEIPTrun_7f2c…19a
COMPLETED

AgentCodex

Protocolapp-server v2

Workspacenorthstar-demo

Verificationpassed

09:41:02 run started

09:41:04 plan updated

09:41:07 approval requested → approved

09:41:13 1 file changed

09:41:16 verification recorded

Git: clean → 1 changedpython3 -m unittest -q

05 / HONEST COMPATIBILITY

Same room.
Different instruments.

AdapterNew runResumeApprovalsPlansUsageTransport
Claude CodeVerifiedVerifiedNot exposedNot exposedNot exposedstream-json
CodexVerifiedVerifiedVerifiedVerifiedVerifiedapp-server v2
OpenCodeVerifiedVerifiedACP modeACP modeACP modeACP / JSON fallback

“Verified” refers to the repository’s pinned protocol fixtures and compatibility tests—not every live CLI version, model, provider, or operating system. The app shows the current local result.

06 / REPRODUCIBLE DEMO

One safe task.
Three isolated runs.

The Northstar fixture has no dependencies, credentials, or network calls. A reset command creates a clean nested Git repository for each agent, so edits and sessions never leak between comparisons.

terminal
make demo-prepare
make demo-verify
Read the demo contract
Short redacted SwiftAgent walkthrough showing agent detection, a run timeline, and a Local Run Receipt
Protocol-fixture walkthrough. No provider or model call; no personal path or credential. Reduced-motion preferences receive the static receipt.

07 / SAFETY WITHOUT THEATER

Two layers.
Both visible.

Native agent approvals and sandbox controls belong to the adapter. SwiftAgent’s own process isolation is a separate layer. The receipt never combines them into a vague “safe” badge.

strict requires usable Linux bwrap and fails closed. fallback is an explicit trusted-machine choice with no OS isolation.

Read the safety notes

08 / FIRST RUN

Bring the agents you already use.

Authentication remains owned by each local CLI. SwiftAgent stores its control-plane data locally and does not ask you to paste provider credentials.

terminal
gh release download v0.6.0 --repo OthmaneBlial/swiftagent \
  --pattern 'swiftagent-v0.6.0.tar.gz' --pattern 'SHA256SUMS'
shasum -a 256 -c SHA256SUMS --ignore-missing
tar -xzf swiftagent-v0.6.0.tar.gz
cd swiftagent-v0.6.0
make install-release
make start-release

Verify the published GitHub attestation too, then open http://127.0.0.1:8000, inspect Your agents, and choose a dedicated workspace.

09 / OPT-IN EVIDENCE

Tell us where the workflow bends.

SwiftAgent uses no silent telemetry. Useful growth evidence is a redacted compatibility report or a clear account of installation, detection, first-run, review, or handoff friction—not a star counter.

10 / SMALL BY DESIGN

InterfaceReact + Vite

Agent-aware composer, execution timeline, receipts, handoffs, files, and settings.

Control planeFastAPI

REST, WebSocket events, lifecycle limits, readiness, and diagnostics.

EvidenceSQLite + Git

Local task history, normalized events, receipts, and workspace change evidence.

ExecutionAdapter registry

Native protocols stay behind a versioned capability contract.